Toward zero-calibration MEG brain-computer interfaces based on event-related fields

Abstract

Magnetoencephalography (MEG) offers high spatiotemporal resolution, but its application in practical brain-computer interface (BCI) systems remains limited partially due to the need for user-specific calibration and inter-subject variability. We present a zero-calibration MEG-based BCI based on event-related fields (ERFs) by leveraging spatial filters and deep learning techniques. First, we developed an on-line ERF-based MEG BCI with a visual oddball paradigm, achieving the mean classification accuracy of 94.29% and an information transfer rate (ITR) of 20.47 bits/min. Using the resulting multi-subject dataset, we applied xDAWN spatial filtering and trained a deep convolutional neural network (DeepConvNet) to classify target versus non-target responses. To simulate real-world plug-and-play use, zero-calibration performance was evaluated using a leave-one-subject-out (LOSO) cross-validation approach. The combination of xDAWN and DeepConvNet achieved the average accuracy of 80.32% and ITR of 12.75 bits/min, respectively, demonstrating cross-subject generalization. These results underscore the feasibility of zero-calibration MEG BCIs for more practical use.