"Addressing between-session variability in MI-BCIs via Optimal Transport"

VICTORIA PETERSON; DIEGO H. MILONE; WYSER, DOMINIK; LAMBERCY, OLIVIER; GASSERT, ROGER; RUBEN SPIES

Virtual

Congreso; 8th International BCI meeting; 2021

BCI Society

Introduction: Brain-computer interfaces (BCIs) based on electroencephalography (EEG) suffer from high signal variability. As a direct consequence, the performance of a BCI system can strongly differ across subjects and sessions. Between-session variability, due to changes at the level of the subject or the electrode position, leads to the need for re-calibration of the whole system before every use. One approach to tackle data distribution drifts is to use optimal transport for domain adaptation (OTDA), in which a non-linear mapping is learned to make the distribution of the calibration samples “similar” to the testing distribution. Then, a new classifier is trained from the transformed calibration samples. In this work, we propose a backward formulation of OTDA in which the mapping is directly learned from testing to calibration data, avoiding classifier retraining.Material, Methods and Results: We evaluated different OTDA alternatives in a 10 subjects motor imagery dataset with two classes (grasping movement vs. relax), acquired in two sessions (S1 and S2). In each session, a total of 160 trials (80 per class) were performed. We simulated a real adaptive scenario by taking S1 as the calibration set. As decoding algorithm, the traditional Common Spatial Pattern (CSP, with EEG patterns filtered between 5 and 30 Hz, 3 pairs of spatial filters) and a Linear Discriminant Analysis classifier were used. OTDA was applied at the feature space level. Each simulated online testing run comprised 20 trials coming from S2. To learn the mapping, a subset of testing data, called here transportation set, was used. This set (with label information), was built using the trials prior to the current testing run (). For both, the original (forward) as well our backward OTDA formulation, the regularized discrete version of optimal transport (OT-S) and its group-sparse version (OT-GL) were implemented. For benchmark comparisons, we also evaluated the performance of a standard calibration(SC) method based on CSP without transfer learning, as well as with the adaptive method proposed in [5], named here as standard recalibration (SR). The overall accuracy across-subjects showed that our proposed backward retraining-free OTDA alternative can provide the same global performance as the retraining SR approach. In addition, our approach is about ten times faster than SR.Discussion: The proposed backward OTDA framework is a retraining-free model which is able to produce equivalent classification results as a complete re-training scheme. The simulated online testing scenario shows that OTDA could be a valuable alternative for rapid multisession BCI use.Significance: Cross-session transfer learning can be conducted by domain adaptation based on optimal transport, in which the data distribution drifts between testing and calibration domains are learned.