A Plug-and-Play P300-Based BCI With Zero-Training Application

Abstract

The practical deployment of P300-based brain-computer interfaces (BCIs) has long been hindered by the need for user-specific calibration and multiple stimulus repetitions. In this study, we build and validate a plug-and-play, zero-training P300 BCI system that operates in a single-trial setting using a pre-trained xDAWN spatial filter and a deep convolutional neural network. Without any subject-specific adaptation, participants could control an IoT device via the BCI system in real time, with decoding accuracy reaching 85.2% comparable to the offline benchmark of 87.8%, demonstrating the feasibility of realizing a plug-and-play BCI. Offline analyses revealed that a small set of parietal and occipital electrodes contributed most to decoding performance, supporting the viability of low-density, high-accuracy BCI configurations. A data sufficiency simulation provided quantitative guidelines for pre-training dataset size, and an error trial analysis showed that both stimulus timing and preparatory attentional state influenced real-time decoding performance. Together, these results demonstrate the real-time validation of a fully pre-trained, zero-training P300 BCI operating on a single-trial basis, without stimulus repetition or user-specific calibration, and offer practical insights for developing scalable, robust, and user-friendly BCI systems.