Towards subject-centered co-adaptive brain-computer interfaces based on backward optimal transport.

Abstract

Objective. Controlling a motor imagery brain-computer interface (MI-BCI) can be challenging, requiring several sessions of practice. Electroencephalography (EEG)-based BCIs are particularly affected by cross-session variability. In this scenario, it is crucial to implement co-adaptive systems, where the machine adapts the decoding algorithm while the user learns how to control the BCI. To support the user learning process, it is essential to measure and provide real-time feedback on self-modulation skills. This study aims to develop a method for online assessment of MI modulation capability to build co-adaptive BCIs that improve both user performance and system accuracy.Approach. Backward optimal transport for domain adaptation allows across-session MI-BCI usage without classifier retraining. Using the cued label to guide the adaptation, a supportive backward adaptation (SBA) method is defined. The required model effort to perform a trial adaptation is proposed as an online metric of MI modulation skills. We conducted experiments on both real and simulated data to demonstrate that this metric effectively informs about the the discriminability and stability of the EEG patterns related to the MI task. The proposed metric is validated by means of Riemannian distinctiveness metrics.Main Results. Our findings show that the associated effort when applying SBA provides a meaningful way of evaluating EEG patterns discriminability, being significantly correlated with Riemannian distinctiveness metrics.Significance. This study introduces a novel framework for co-adaptive BCI learning that performs data adaptation while assessing the MI-BCI skills of the user. The proposed SBA approach can enhance BCI performance by facilitating session-to-session adaptation and empowering users with valuable feedback based on their current MI modulation strategy. This framework represents a significant advancement in developing user-centered, co-adaptive MI-BCIs that effectively support and enhance user capabilities.