Artifact (artefact) reduction in EEG – and a bit of ERP basics CNC, 19 November 2014 Jakob Heinzle Translational Neuromodeling Unit Overview • Basic Principles of ERP recording (Luck Chapter 3) • Averaging, Artifact Rejection and Artefact Correction (Chapter 4) • A multiple source approach to the correction of eye artifacts (Berg and Scherg, 1994) EEG artefacts 2 Hansen’s axiom • “There is no substitute for good data!” • Get your data “free of noise” during recording already. – No electromagnetic contamination (Faraday cages, no screens inside etc.) – No eye movements, no muscle artifacts, no sweating (Instruct subjects and make it comfortable for them.) – No bridging etc. (careful setup of caps etc.) EEG artefacts 3 Basics of ERP (EEG) recording • Electrodes (Ground and Reference) – Often Mastoid reference (average over both mastoids) – Signal is A – (Lm/2 + Rm/2), where all A, Lm and Rm are voltages relative to ground. – Sometimes average reference. • Typical size of ERP is about 10 mV EEG artefacts 4 EEG electrodes EEG artefacts 5 Sources of noise • Everything that can cause a voltage difference between two electrodes and is not of “brain origin” EEG artefacts 6 Environmental noise • Electrical noise in the environment – power line AC (50 Hz), Video monitors (refresh rate), Impedance changes at electrodes, bridges, … • Reduce noise as much as possible – Faraday cages, shielded room, etc. – Reduce impedance at electrodes (gel, scratch surface of skin, …) EEG artefacts 7 Amplification, Filtering and Digitization • Active amplifiers increase signal to range that is then digitized into 4096 (212) discrete steps. – Set gain of amplifier to use entire range • High pass filtering of signal (often 0.01 Hz) • Sampling rate depends on low pass filter of amplifier Nyquist. EEG artefacts 8 Averaging • In most cases ERP signals are averaged. – Assumptions: Signal always the same and only EEG noise varies from trial to trial. – If noise is independent of ERP it is reduced by a factor 1/sqrt(n) • “It is usually much easier to improve the quality of your data by decreasing sources of noise than by increasing the number of trials.” EEG artefacts 9 Averaging EEG artefacts 10 Latency variability EEG artefacts 11 Overlap between trials Problematic if different for different trial types. EEG artefacts 12 Averaging • Area measures are less sensitive to latency variability. • Response locked averaging. • Woody filter. Iterative template matching, template calculation technique. • Time locked spectral averaging. EEG artefacts 13 Time locked spectral averaging EEG artefacts 14 Steady state ERP Use overlap and drive responses into a steady state. EEG artefacts 15 Typical artefacts from participant • Eye blinks • Eye movements • Muscle activity • Skin potentials • Heart artefacts • … All of those can create large signals and might be correlated with the task. EEG artefacts 16 Some examples EEG artefacts 17 How to deal with artefacts • Artefact rejection: Remove all trials that contain contaminated data. • Artefact correction: Use all data, but try to correct for the artefacts. • But, best thing is always to avoid artefacts as much as possible. EEG artefacts 18 Post-processing of artefacts • Detecting artefacts is a signal detection problem. • Problem: Threshold for artefact detection. Typical ROC type problem (True positive vs. false positive) In general: Define artifact measure, detect artifacts, reject artifacts. EEG artefacts 19 Electric field of the eyes http://www.bem.fi/book/28/28.htm EEG artefacts 20 Example: Blinks EEG artefacts 21 Eye movement artifact correction EEG artefacts 22 Basic idea – component model • EEG data is modeled as sum of EEG and eye artefact components. • Spatial distribution (scalp distribution) activated by a temporally evolving factor. EEG artefacts 23 What are the components? • Eye components are derived from a calibration session prior to the experiment. – Eye movements into different directions and blinks (every 2 secs). – PCA on this data: 3 components explain 95% of variance. • EEG components are fitted dipole sources, or combination of assumed dipoles. – No details here, different paper of the authors. EEG artefacts 24 Different models EEG artefacts 25 Eye movement results EEG artefacts 26 Eye movement results EEG artefacts 27 Testing the method Use “artefact free” data and data with artefacts. For both compare optimizing (dipole fitting), surrogate and traditional method. EEG artefacts 28 fMRI results – Visuomotor mismatch specific activation EEG artefacts 29 EEG artefacts 30 Residual variance in individual subjects EEG artefacts 31 Results - Maps EEG artefacts 32 Results - Maps EEG artefacts 33 Spatial accuracy (consistency) Compared to uncorrected model without EOG electrodes. EEG artefacts 34 Results • Optimized methods seems to be best • Artefact rejection does not remove all eye movement artefacts. • Ground truth is not known, but they take one of the fitted results to compare. EEG artefacts 35 ICA based artefact removal • Independent component analysis (ICA) can be used to find independent sources and exclude sources that come from artifacts. 𝑥 𝑡 = 𝐴 ∙ 𝑠(𝑡) • ICA assumes x(t) is a linear mixture of (maximally) independent sources. • For details see e.g.: – ICA general: Hyvärinen and Oja, Neural Networks, 13(4-5):411-430, 2000 – ICA in EEG: Delorme et al, IEEE 2005 and many other papers from Scott Makeig’s group. EEG artefacts 36 Some more sources • Some EEG artifacts reviewed: – https://www.youtube.com/watch?v=1LftSdvNXh0 • Web based EEG Atlas – http://eeg.neurophysiology.ca • Saccadic spike artefact in MEG – Carl et al, Neuroimage 59:1657 2012 EEG artefacts 37