ARfitStudio

From SCCN
Jump to: navigation, search

Contents

What is ARfitStudio

First of all, ARfit is a collection of Matlab modules developed by Tapio Schneider and Arnold Neumaier for estimating parameters of multivariate autoregressive (AR) models, diagnostic checking of fitted AR models, and analyzing eigenmodes of fitted AR models. This plugin, ARfitStudio, uses it so that one can interactively clean TMS-induced short-burst (5-10 ms) artifacts and so on. The nature of the artifact does not matter (could be spikes by artifact, could be zeros, NaNs, etc), as long as the artifact duration is short and they are correctly marked. It has strength on the following points.

  • To perform quick (e.g. smart epoching & integration to continuous data using mcolon(); before and after correction comparison with one click, etc) and intuitive (e.g. training, correction, and blending windows overlaid on grand-median ERPs) correction of spiky artifacts.
  • Immediately usable after importing the continuous data; it is the very first stage of the preprocess pipeline.

This plugin is dependent on ARfit and mcolon. ARfit can be installed by installing SIFT from EEGLAB plugin manager. To use mcolon(), users should compile the mex file.

http://www.mathworks.com/matlabcentral/fileexchange/174-arfit
http://www.mathworks.com/matlabcentral/fileexchange/29854-multiple-colon/content/mcolonFolder/mcolon.m

Why ARfitStudio

If there are spike artifacts with high amplitude, we cannot filter the data (because spikes spread in the time domain). In certain situation, it is more important to make them harmless for the sake of data process, rather than recovering the underlying signals (the latter is often impossible). This plugin provides a solution for it.

How it works

It learns autoregressive model from the training period (green), which is the past of the spike, to make a prediction to replace problematic spike data points (red). For smooth connection, one can also set optional blending period (light magenta) during which predicted signal and original signals are gradually mixed with a linear slope. Thus, this plugin replaces the bad data points using the past information (therefore temporal interpolation). It does NOT recover the signal under the noise. One can select multiple markers for correction. In the main plot, grand-median ERPs for all the channels are shown. To save the result, just save the EEG dataset (do not check 'restore the last data' for this!)


Screenshots

Figure 1. interpolateSpike: GUI

Figure 2. interpolateSpike: Before correction

Figure 3. interpolateSpike: After correction

Data by courtesy of Michael Borich


If you want to use ICA for the final analysis

Because this ARfit-based interpolation is performed for the same time points across channels, ICA cannot decompose this period. As a result, you'll see strange spikes in IC activations (i.e., EEG.icaact) during the correction windows. In the plots below, I show 1) ERP of IC activation for 33ch with ARfitStudio applied on channels (left), 2) the same but ARfitStudio applied on ICs (right). The noise was originally present within -/+10 ms relative to latency zero.

OnChannels.png OnICs.png

Performing ARfit-interpolation for ICA properly is complicated.

  1. Import data.
  2. Perform ARfitStudio on channels.
  3. Preprocess the data (high-pass filter, CleanLine plugin, clean_rawdata plugin, average referencing, AMICA plugin, DIPFIT plugin, fitTwoDipoles plugin). Note that you need to exclude the interpolated data from ICA process (those have no multivariate property across channels, so not decomposable).
  4. Perform ARfitStudio on IC activations.
  5. Backproject interpolated IC activation to reconstruct channel data.

Note that the ARfitStudio is used twice, and the first application is purposed only for various filters, not even for ICA. In this way, you can obtain both clean channel signals and IC activations. For detail, see 'batchDemoForICA.m' included in the download package. I recommend you take a look because there are several cumbersome steps you have to go through.

Note for batch users (03/11/2019 updated)

If one wants to run the process as a batch,

  1. Use EEGLAB function pop_epoch() to crop out peri-event data points. Note all the negative latencies will be used as a learning period for ARfit.
  2. Use arfit2interpolate(). Note that the last input 'last_n_pointsToBlend' does NOT specify the additional length to the data points to the main interpolation window (as the GUI operation indicates), but it actually specifies the last data points of the main interpolation window. For example, outputData = arfit2interpolate(EEG.data, [80 95], 5) means, [80:90] for 100% interpolation, and [91:95] is blended with [83% 67% 50% 33% 17%] interpolated data blended with [17% 33% 50% 67% 83%] of the original data, respectively.
  3. Use putBackEpoch2Continuous() to re-construct the corrected, epoched data to the continuous data.


Reference

A. Neumaier and T. Schneider, 2001: Estimation of parameters and eigenmodes of multivariate autoregressive models. ACM Trans. Math. Softw., 27, 27-57.

T. Schneider and A. Neumaier, 2001: Algorithm 808: ARfit – A Matlab package for the estimation of parameters and eigenmodes of multivariate autoregressive models. ACM Trans. Math. Softw., 27, 58-65.

Authors: Makoto Miyakoshi and Tim Mullen. SCCN, INC, UCSD