Looking at ERPs from Another Perspective: Polynomial Feature Analysis

Zusammenfassung (Abstract) :

Event-related potentials (ERPs) are classically studied measuring amplitude and latency characteristics of individual components. Such analysis is restricted to individual time points and largely ignores the time-series nature of the ERP. This motivates alternative preprocessing algorithms that might reveal new information about the signal decoded in the temporal relationships between neighbouring data points. In the current work, we applied polynomial fits of orders one to four to ERPs (average and individual epochs) before analyzing the signal. Depending on other pre-processing methods (like subsampling and filtering), a low order polynomial should, in principle, be able to capture the ERP shape and reduce noise in single-trials. The polynomial fits were performed on individual ERP segments and the analysis was performed with the corresponding coefficients instead of the amplitude values. For the analysis we used data from an oddball task evoking a broad P300 component (five subjects, two sessions each). The best combination of coefficients was derived from the performance of a support-vector machine (SVM) classifying ERPs labelled as "standard" and "target", respectively. The corresponding ERP topographies (both, average and single epochs) strengthen the notion that analysis of polynomial features provides a tool for exploration of new relationships in ERP data.

Stichworte :

ERP, single-trial, polynom, SVM, oddball, P300