Co-adaptive BCI based on supervised domain adaptation: results in motor imagery simulated data

VALERIA SPAGNOLO; CATALINA MARÍA GALVÁN; NICOLÁS NIETO; DIEGO H. MILONE; RUBÉN DANIEL SPIES; VICTORIA PETERSON

Bruselas

Congreso; 10th International Brain-Computer Interface Meeting 2023; 2023

BCI Society

Introduction: Brain-computer interfaces (BCIs) can be thought of as a two-learners system, in which theuser learns how to control the computer and, simultaneously, the computer learns how to decode theuser’s brain activity [1]. When used across several sessions, as in motor imagery (MI) BCIs for rehabilitation,the recorded EEG signals contain high variability. Machine learning systems used to decode brain activityshould then adapt to those signal changes and help the user in the development of stable EEG patterns. Inthis line of work, a backward formulation of optimal transport for domain adaptation (BOTDA) wasproposed in [2] to avoid recalibration in cross-session MI-BCIs. Although BOTDA showed promising resultsin a supervised sample-wise scenario, it remains to be elucidated whether the success of the adaptationdepends on the subject's ability to perform the MI task or on the adaptive capabilities of the model. Herewe hypothesize that supervised adaptation based on BOTDA is successful only when: H1) the EEG patternsprovided by the user corresponds to the mental task to be performed and H2) the calibration data, usedto train the decoding model, is discriminative enough from the decoding system viewpoint.Material, Methods and Results: Considering MI-BCIs for motor rehabilitation, realistic MI vs. Rest EEG datawas generated based on a custom implementation that extends [3]. MI alpha desynchronization (aka ERD)was simulated in the left hemisphere for MI. “Rest” corresponded to no-ERD. We used the first session(S1) to train the model (calibration data) and the following session (S2) was used as testing data. For eachsession, 100 trials of each class were simulated. As a decoding algorithm, a common spatial pattern and alinear discriminant analysis were used, as in [4]. To prove H1 we simulated S1 as the ideal case, i.e the ratiobetween ERD and baseline amplitudes (%ERD) was set to 50 for all trials belonging to the MI class (0% offailed MI trials). Results show that BOTDA can provide almost perfect classification accuracy (ACC)regardless of the %ERD in S2 (e.g. ACC=0.97 with BOTDA, ACC=0.51 without BOTDA for a S2 with%ERD=10). Experiments varying the percentage of failed MI trials, but with high %ERD, indicated thatBOTDA could not help with failed trials (e.g. ACC=0.76 with BOTDA, ACC=0.76 without BOTDA for a S2 with%ERD=45 and 50% of failed MI trials). To validate H2 we trained the decoding model with data fromsessions with different %ERD values. We found that when the simulated calibration data did not containdiscriminable ERD patterns (%ERD