Minimizing the cross-session variability in MI-BCI with Optimal Transport

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

Simposio; CYBATHLON Symposium 2020; 2020

Brain-Computer Interfaces (BCIs) based on electroencephalography (EEG) transform electrical brain activity into output commands. The decoding algorithms are learned using training data (calibration set). If the distribution of new unseen data differs from calibration, the performance of the decoding algorithm may be poor. This is a direct consequence of EEG variability, which can be due to changes at the level of the BCI user or the electrodes positions. In order to avoid the lengthy re-calibration step of the decoding algorithm before every BCI session, we propose to use Optimal Transport for Domain Adaptation (OTDA) [1] in a novel backward formulation that, unlike the original one, allows learning the transport from the new session to the previous one, avoiding classifier retraining. Considering real-time constrains, we propose to learn the transportation plan using a subset of testing data, called here transportation set. We evaluated the performance of the proposed method in a 10 subject Motor Imagery (two classes: grasping movement vs. relax) dataset acquired in two sessions (S1 and S2, respectively) [2]. In each session a total of 160 trials (80 per class) were performed. We took S1 as the calibration set and S2 as the testing set. As decoding algorithm, the traditional Common Spatial Pattern (CSP) and a Linear Discriminant Analysis (LDA) framework were used (EEG patterns filtered between 5 and 30 Hz, 3 pairs of spatial filters). OTDA was applied at the feature space level. The overall accuracy across-subjects showed that the proposed backward OTDA method yielded overall classification improvements of up to 15% as compared to CSP+LDA without transfer learning, when only 20 trials of a new session were used to make the adaptation. These results show that intra-session variability can be minimized by means of OTDA, reducing the number of recalibration samples and increasing classification accuracy.