Signal-Adaptive Denoising Of Event-Related Potentials

In the present study we propose a data-driven, fully unsupervised denoising approach for multi-trial univariate signals. The proposed methodology is based on Empirical Mode Decomposition (EMD) and hence also applicable for transient or non-stationary signals. The rationale of the presented method is that different realizations (multiple trials) of the same underlying process have also similar intrinsic signal components. These components may be extracted by EMD for each single realization and finally, the entirety of all signal components forms clusters of corresponding components with similar spectral characteristics. A denoising is then tantamount to identifying the cluster(s) containing high-frequency noise components. The effectiveness of the proposed methodology is demonstrated on the basis of visual event-related potentials (ERPs) of dyslexic and normal control children. We could show that the novel method allows for a reliable ERP estimation and that it provides a tool for an objective extraction of ERPs on both a single-subject as well as on a single-trial basis.