About EEG Signal Cleaning

From Nemo

Jump to: navigation, search

Project Summary

EEG data are a time series of scalp electrical potential from 1 or more recording sites or channels. This electrical potential, recorded as EEG, is generated by the superposition of electrical activity from both cortical and non-cortical sources. Signal cleaning refers to the process of extracting the EEG's non-cortical signal components with minimal distortion to the remaining cortical signal.


Point Person: Robert Frank.
Other personnel: Gwen Frishkoff, Cathy Poulsen, Chris Hoge, Matt Sottile.

Project Description

EEG data (and derived ERPs) often contain unwanted signals (“noise”), generated by noncerebral sources (eye blinks, EKG, muscle movement, etc.), as well as by sources outside the body, such as electrical equipment that is present during the time of recording (resulting in 50 or 60hz “machine noise”). The goal of our EEG signal cleaning project is to design, implement, and evaluate various methods for identifying and removing these unwanted signals. Although some signal cleaning methods (e.g., the Gratton method) can be applied to averaged EEG (i.e., ERP) data, the methods that appear to be most effective (e.g., ICA and other blind source separation algorithms) are typically applied to the raw EEG. Our current EEG signal cleaning tasks can be divided into 3 subcategories: (a) Design & implementation of algorithms; (b) Tuning and optimization of algorithms for identification and removal of different types of artifacts (EOG, EKG, EMG, etc.), and (c) Tools for rigorous evaluation of signal cleaning (assessment of whether cortical signals were left intact and whether targeted “noise” was completely removed).

Design & implementation of algorithms
Methods for EEG signal cleaning fall into three broad categories: blind component analysis (ICA, PCA, SOBI); regression (Gratton method; cf. more recent implemetations); and a newly developed approach we call directed component analysis, or DCA. Our NIC signal cleaning toolbox currently includes implementation of the following methods:… [RMF, Chris?]. In addition, the NIC has developed parallelized versions of two ICA methods, Infomax and Fast-ICA. These high-performance implementations are available through the NIC’s HiPerSAT toolbox.
Optimization for different types of artifacts
Different artifacts have different signal characteristics (frequency, amplitude, topographic distribution), which may call for customization of signal cleaning methods. For example, in a previous evaluation study, we found that Infomax outperformed both FastICA and SOBI with respect to the identification and removal of eye blinks (Frank & Frishkoff, 2007). By contrast, SOBI outperformed Infomax when both methods were used to separate EKG (cardiac signal) from interictal spiking in several epileptic datasets. One implication is we need sophisticated methods of evaluating EEG signal cleaning outcomes, to determine whether a particular outcome is satisfactory, where “satisfactory” may be defined differently in different task contexts.
Tools for routine evaluation of signal cleaning outcomes
We have designed a framework called APECS (automated protocol for evaluation of electromagnetic component separation) for routine evaluation of signal cleaning outcomes. APECS is not designed with the goal of determining which algorithm is “best,” independent of specific applications. Rather, we assume that every method will sometimes succeed and sometimes fail. Characteristics of the test data and of the targeted artifacts, the algorithms used for signal cleaning, and their specific implementation can all affect signal cleaning outcomes. Further, success or failure of a method is not typically all-or-none. In a given dataset, some instances of a targeted artifact may be completely identified and removed, while other instances may be missed or only partially separated from the brain signals of interest. To have confidence in the application of a signal cleaning method to a given dataset, we feel that evaluation should be routinely performed.
/wiki/images/1/17/Fish1.png /wiki/images/e/ea/Fish2.png /wiki/images/f/fa/Fish3.png /wiki/images/f/ff/Fish4.png /wiki/images/4/40/Fish5.png /wiki/images/c/c5/Fish6.png
Personal tools