Evaluating Performance of EEG-based Brain-Computer Interface Speller with Dynamic Stopping Criterion

A P300 Event-Related Potential (ERP)-based brain–computer interface (BCI) speller system offers a noninvasive means of communication by identifying target ERPs from non-target ERPs in EEG signals. Conventional evaluation of prediction performance only tracked prediction accuracy and relied on spelling the keys with a fixed but large sequence replications, which ignored the spelling speed and overall efficiency. This paper presented a comprehensive analysis of model performance with a dynamic stopping strategy that adaptively terminated the flash sequence once character-level classification confidence exceeded a predefined threshold. The approach was first validated through simulations and then applied to real EEG data from 23 participants using a Stepwise Linear Discriminant Analysis (SWLDA) classifier under the Row-Column paradigm. Stop thresholds ranging from 0.6 to 1.0 were evaluated. Information transfer rate (ITR) and BCI Utility were calculated. We identified three clusters of participants. Results of real data analysis demonstrated that implementing an appropriately selected stop sign threshold could improve spelling efficiency, offering a practical enhancement to real-time BCI speller performance.