Empirical multi-scale thresholding for evoked neural activity denoising

Evoked potentials (EPs) are responses elicited by stimulation of the nervous system that serve as key biomarkers for assessing neural function, connectivity, and pathophysiology. Reliable EP extraction is challenged by low signal amplitudes, unrelated neural activity, and background noise across overlapping frequency ranges.

This study presents a novel framework to estimate the noise distribution around EPs without relying on prior assumptions. The method uses a multi-scale bootstrap approach to statistically characterize noise and uncertainty, allowing separation of meaningful EP components from unrelated background activity. The core principle of the bootstrap is that the variance of resampled distributions empirically estimates variability, enabling noise characterization around the mean. By applying this strategy across multiple frequency bands, the method effectively captures dynamic neural variations and improves EP detection reliability.

The method is evaluated using electrocorticographic (ECoG) recordings, including synthetic and real EPs. Quantitative analysis showed lower mean square error (MSE) between denoised and true EPs, indicating improved signal-to-noise ratio (SNR). Qualitative evaluation of real EPs demonstrated enhanced visualization and more accurate morphology recovery, with reduced false detections and preserved EP integrity.

Compared with conventional filtering techniques, the proposed method better adapts to non-stationary noise and dispersed EP energy while maintaining computational efficiency, ease of implementation, and adjustable confidence levels.

This approach offers improved EP detection and visualization in clinical and research contexts, particularly where recordings are time-limited or patient tolerance for extended sessions is low, supporting broader applications in neuroscience and neuro-engineering.