Exploring new territory: Calibration-free decoding for c-VEP BCI

TL;DR

This study explores two zero-training methods aimed at enhancing the usability of brain-computer interfaces (BCIs) by eliminating the need for a calibration session, and compares UMM to the state-of-the-art c-VEP zero-training method that uses canonical correlation analysis (CCA).

Abstract

This study explores two zero-training methods aimed at enhancing the usability of brain-computer interfaces (BCIs) by eliminating the need for a calibration session. We introduce a novel method rooted in the event-related potential (ERP) domain, unsupervised mean maximization (UMM), to the fast code-modulated visual evoked potential (c-VEP) stimulus protocol. We compare UMM to the state-of-the-art c-VEP zero-training method that uses canonical correlation analysis (CCA). The comparison includes instantaneous classification and classification with cumulative learning from previously classified trials for both CCA and UMM. Our study shows the effectiveness of both methods in navigating the complexities of a c-VEP dataset, highlighting their differences and distinct strengths. This research not only provides insights into the practical implementation of calibration-free BCI methods but also paves the way for further exploration and refinement. Ultimately, the fusion of CCA and UMM holds promise for enhancing the accessibility and usability of BCI systems across various application domains and a multitude of stimulus protocols.

Figures (2)

• Figure 1: Bandpass hyper-parameters for CCA and UMM. Depicted are the grand average classification accuracy for CCA and UMM across varying highpass (left) and lowpass (right) cutoff values. Here, a single-trial duration of 31.5 s is used. When varying the highpass, the lowpass remained at 40 Hz, and when varying the lowpass, the highpass remained at 6 Hz. In order, the symbols behind a method refer to: the type of covariance matrix being empirical (e) or block-Toeplitz (t); covariance matrices computed instantaneously (1) or cumulative (c); and the mean vectors (of UMM) computed either instantaneously (1) or using a weighted cumulative average (w). The dashed gray line denotes the theoretical chance level (5%).
• Figure 2: Decoding curve for CCA and UMM. Depicted are the grand average classification accuracy for CCA and UMM across varying single-trial durations. Here, a bandpass of 6-50 Hz is used. For a definition of method names, see \ref{['fig:decodingcurve']}. The dashed gray line denotes the theoretical chance level (5%).