Event-related potential extraction workflow based on kernel density estimation

Event-related potentials (ERPs) are a critical neuroscientific tool for investigating brain responses to external stimuli and serve as a key linking mechanism in brain–computer interface systems. Traditional ERP extraction methods rely on threshold-based trial rejection and time-locked averaging techniques, which often have limited ability to handle outlier data and are susceptible to random artifacts. To address this, we propose a novel ERP extraction workflow based on kernel density estimation. We construct trial-wise datasets at the sampling-point granularity and model the probability distribution of each trial using Gaussian kernel density estimation, effectively reducing outlier influence while preserving all trial data. The fitted probability density function serves as the objective function for ERP extraction, enabling active reconstruction of optimal ERP waveforms by incorporating inherent EEG temporal dependencies. Specifically targeting uneven noise distribution across multiple channels, we introduce an adaptively steering kernel dynamically generated from electrode covariance matrices, which optimizes the adaptive matching of inter-channel noise structures to ensure more precise density function fitting. Using two real datasets and simulated datasets, our comparative analyses of baseline root mean square error, component-level statistical metrics, and residual correlations demonstrate that, compared with the traditional trial rejection and time-locked averaging methods, our approach exhibits outstanding effectiveness in isolating ERP components from raw signals and significantly reduces the impact of outlier contamination.