Tangent Space Features-Based Transfer Learning Classification Model for Two-Class Motor Imagery Brain–Computer Interface

Abstract

The performance of a brain–computer interface (BCI) will generally improve by increasing the volume of training data on which it is trained. However, a classifier’s generalization ability is often negatively affected when highly non-stationary data are collected across both sessions and subjects. The aim of this work is to reduce the long calibration time in BCI systems by proposing a transfer learning model which can be used for evaluating unseen single trials for a subject without the need for training session data. A method is proposed which combines a generalization of the previously proposed subject-specific “multivariate empirical-mode decomposition” preprocessing technique by taking a fixed band of 8–30[Formula: see text]Hz for all four motor imagery tasks and a novel classification model which exploits the structure of tangent space features drawn from the Riemannian geometry framework, that is shared among the training data of multiple sessions and subjects. Results demonstrate comparable performance improvement across multiple subjects without subject-specific calibration, when compared with other state-of-the-art techniques.

Deep Learning Based Inter-subject Continuous Decoding of Motor Imagery for Practical Brain-Computer Interfaces

“…Previous literature on inter-subject transfer learning using CSP yielded the best average accuracy of 79% on BCI Competition III, dataset IVa, where the number of subjects was 5 (Devlaminck et al, 2011 ). Tangent space features drawn from the Riemannian geometry framework were used for transfer learning using BCI competition IV, dataset 2a, which achieved an average leave-one-subject-out-cross-validation accuracy of 75.52% (Gaur et al, 2019a ). In a recent study, Halme and Parkkonen reported inter-subject transfer learning accuracy in EEG of 67.7% on their own experimental data using CSP with logistic regression (Halme and Parkkonen, 2018 ).…”

“…In a recent study, Halme and Parkkonen reported inter-subject transfer learning accuracy in EEG of 67.7% on their own experimental data using CSP with logistic regression (Halme and Parkkonen, 2018 ). Although a direct comparison is not possible here as the datasets used in Gaur et al ( 2019a ), Halme and Parkkonen ( 2018 ), and Devlaminck et al ( 2011 ) were different but the average of 7 out of 9 subjects [ignoring subject 2 and subject 3 due to poor data quality as revealed by BCI competition results (BCI-Competition, 2008 )] in our case achieved an average single-trial classification accuracy close to 75% (Adam: 75.13% and SGDM: 74.43%). It is noteworthy that previous studies on inter-subject transfer learning mentioned above did not deal with continuous decoding and used traditional approaches rather than deep learning.…”

“…An extensive detail of transfer learning approaches for BCIs has been given in Jayaram et al (2016). Transfer learning is often implemented by transferring stationary and/or discriminative information invariant across the subjects (Wang et al, 2015;Gaur et al, 2019a). Apart from globally relevant feature representation, other approaches to transfer learning involve ensemble learning, sparse subset of spatial filters, and classifiers (Fazli et al, 2009;Tu and Sun, 2012;Raza et al, 2019), and domain adaptation of classifiers (Vidaurre et al, 2011).…”

“…Some different approaches are also proposed for inter-subject transfer learning where event-related cortical sources are estimated from subject independent EEG recordings (Saha et al, 2019 ) which can compensate for the changes in head morphology and electrode positioning (Wronkiewicz et al, 2015 ). In a recent study, a Riemannian geometry-based approach is successfully applied for cross-subject and cross-session transfer learning which significantly improved BCI performance (Zanini et al, 2018 ; Gaur et al, 2019a ). Others have also used novel filtering techniques using multivariate empirical mode decomposition (MEMD) along with CSP features for subject independent learning and have shown improved performance on BCI Competition IV-2a dataset (Gaur et al, 2019a , b ).…”

“…In a recent study, a Riemannian geometry-based approach is successfully applied for cross-subject and cross-session transfer learning which significantly improved BCI performance (Zanini et al, 2018 ; Gaur et al, 2019a ). Others have also used novel filtering techniques using multivariate empirical mode decomposition (MEMD) along with CSP features for subject independent learning and have shown improved performance on BCI Competition IV-2a dataset (Gaur et al, 2019a , b ). Halme and colleagues compared several different methods for cross-subject decoding of MI and passive movements using both EEG and MEG signals.…”