In general, the preprocessing methods used in EEG are very dependent on the goal of the applications. Having said that, there are some methods that are used very commonly to improve the quality of Signal to Noise ratio, such as Common Average Referencing (CAR) or filtering. It would be interesting to summarize the effective signal preprocessing methods since they usually can be similar in different applications. What are the methods that you have found effective on your data?
Adham provided some suggestions re dimension reduction, but sometimes it is also possible to multiply the data, as in our series of papers such as:
https://www.researchgate.net/publication/230754283_Multiplication_of_EEG_samples_through_Replicating__Biasing__and_Overlapping
You can take existing epochs and use different subsamplings or subperiods as new epochs, and it is even useful for these to be overlapped, or biased to the central and more consistent part of the epoch.
However, you asked specifically about Laplacian, and we have also shown that this is the best method we have found in terms of getting rid of unrecognized muscle (EMG):
https://www.researchgate.net/publication/224866430_Surface_Laplacian_of_Central_Scalp_Electrical_Signals_is_Insensitive_to_Muscle_Contamination
This work is very interesting because the demonstration of the elimination of EMG is based on a unique paralysis dataset in which subjects are paralysed to eliminate it!
Conference Paper Multiplication of EEG Samples through Replicating, Biasing, ...
Article Surface Laplacian of Central Scalp Electrical Signals is Ins...
I usually apply Independent Component Analysis (ICA) on the raw data. Then correlation of Component-Data with EOG-electrodes can be used for an automatic removal of occular artefacts. or visual inspection of the component data helps to identify other artificial components which then can be rejected.
With my experiences, I can said that the impedances must be very lower, the machine must have a independent connection to eletricity. In the ICU you must isolate the eletrodes with some kind of cotton cloth, diaper or cheesecloth, when the noise persist we can use 2 electrodes intertwined stuck in equipament near the EEG box,...
Hello Pouya. I usually remove 0-1 Hz direct current interference with trials. baseline correction, visual artefact inspection in addition to removal algorythms also help. What specifically you woul dlike to know? Best, naira
Hi Naria,
Apart from general approaches that can improve the quality of signal, I am more interested about the spacial filters such as Laplacian spatial filter or other filters designed for a given task.
AC line frequency and its first few harmonics (60,120, 180 Hz in US) have to be notch-filtered out first. Low frequencies which often contain amplifier artifacts (
Thank you for your answers. I have a follow up questions.
About the spacial filters; do they change the number of electrodes to reduce the dimensions?
What are the methods that don't change the number of dimensions and just improve the signal quality of each channel (e.g. bandpass filtering or CAR)?
Pouya, I agree with Matthias. My comment in here is mainly on dimension reduction (DR) part of your previous discussion. First I should mention that Dimension Reduction (DR) is a favorable preprocessing stage that is often used to address the dimensionality problem of EEG data through either feature or electrode reduction. This is because when you are using a classifier, it is desirable to have a smaller number of features*electrodes in compare with what you have as the number of trials or samples.
It is possible to find similarities in terms of principles between CSP and methods like PCA, SVD, and ICA. Essentially, in all of these methods, it is endeavored to maximize or minimize a relation between components of the signal/s (i.e., independency between electrodes in ICA, spatial differences between classes/samples in CSP, separability in SVD, and un-correlatability in PCA).
You might choose to keep the transformed signal the way it is or only keep a partial data that lets say represent 99% of the variance/information (in which case you would have a decomposed signal. With CSP, the disadvantage is that it requires the entire trial duration for calculating the necessary components, which makes it unsuitable for real time EEG-BCI systems. Subject transfer (using CSP components that are constructed using other subjects EEG data or previously sessions of the same subject) is a solution to this problem. The other drawback is that these decomposition type methods use procedures that mostly isolates a set of features or electrodes that satisfy a certain criterion without considering their impact on the classifier's learning rule.
Evolutionary methods such as GA and PSO are alternative approaches used for dimension reduction. These approaches evolve their population toward a subset of features or electrodes that achieved the best classification performance, with the drawback of being computationally intensive and unsuitable for real time systems. This problem can also be solved by shifting the computationally intensive operation of these EA-based methods to off-line mode through subject transfer.
Impedance matching; proper preparation of the patient; sleep deprivation, sleep EEG recording
Adham provided some suggestions re dimension reduction, but sometimes it is also possible to multiply the data, as in our series of papers such as:
https://www.researchgate.net/publication/230754283_Multiplication_of_EEG_samples_through_Replicating__Biasing__and_Overlapping
You can take existing epochs and use different subsamplings or subperiods as new epochs, and it is even useful for these to be overlapped, or biased to the central and more consistent part of the epoch.
However, you asked specifically about Laplacian, and we have also shown that this is the best method we have found in terms of getting rid of unrecognized muscle (EMG):
https://www.researchgate.net/publication/224866430_Surface_Laplacian_of_Central_Scalp_Electrical_Signals_is_Insensitive_to_Muscle_Contamination
This work is very interesting because the demonstration of the elimination of EMG is based on a unique paralysis dataset in which subjects are paralysed to eliminate it!
Conference Paper Multiplication of EEG Samples through Replicating, Biasing, ...
Article Surface Laplacian of Central Scalp Electrical Signals is Ins...
You might want to look into some work by Vernon Lawhern: see "Detection and classification of subject-generated artifacts in EEG signals using autoregressive models". He also put together a MATLAB toolbox that people can use.
Hi. At our laboratory we measure EEG during exercise, for example running and cycling.
We use Active and shielded electrodes which reduces noise a lot. Make sure you use a minumum of 64 channels and equidistant electrode positions that extend below the hat brim line. We also developed cap that holds fixed the cables, so no cable movement artifacts. Then for signal analysis we use only signals from 0.2 Hz to 40Hz. IF we are interested in frequencies above, we make sure that its below the frequencies of movement. Also use a notch filter, 50 or 60Hz. Then we use BSS-CCA to reduce noise from muscles and other sources. Alternatively we also use AMICA and remove noise components. You may also have a look into PARAFAC. About the channels referencing. We prefer local common average because its better for source localisation. This works well :D
It is generally a good idea to combine different approaches, because they work differently well with different kinds of artifacts. Do not rely too much on a single approach!
Some suggestions to improve your SNR
1. use more trials
2. filter your data. For ERPs, a 0.1-30Hz bandpass is fine to get rid of skin potentials (=low frequency bands) and muscle & electrical line noise (=high frequency bands)
3. use thresholds. Amplitude thresholds are a rather crude method, but still useful as "last line of defence" against nasty artifacts that got past all other artifact correction/rejection methods.
4. visually inspect the EEG and remove artifactual epochs. Automatic procedures alone are no guarantee for clean data.
5. [optional] in high-density recordings, interpolate artifactual channels (e.g. by using Faster)
http://www.mee.tcd.ie/neuraleng/Research/Faster
6. [optional] use independent component analyisis (ICA) to get rid of artifactual components, e.g. blinks
http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2895624/
... the list is incomplete, but a start.
cheers
Tobi
PS: Common average is a way to reference data - it is not an artifact correction/rejection procedure.
Tobias,
I have kept my low-frequency cutoffs lower (at 0.01 Hz) for the ERP filtering based on those two articles:
Acunzo DJ, MacKenzie G, van Rossum MCW. 2012. Systematic biases in early ERP and ERF components as a result of high-pass filtering. Journal of Neuroscience Methods 209:212–218. http://dx.doi.org/10.1016/j.jneumeth.2012.06.011.
Widmann A, Schröger E. 2012. Filter effects and filter artifacts in the analysis of electrophysiological data. Frontiers in Perception Science:233. http://dx.doi.org/10.3389/fpsyg.2012.00233.
However I am not an expert on the topic so have anyone tried to vary the low cutoff frequency and do you agree with the recommendation from Acunzo et al. (2012):
"To choose the cut-off of the high-pass filter, we reiterate the guideline stated in Luck (2005): set the high-pass filter cut-off value to 0.01 Hz by default. For less docile participants whose data may be noisier, such as children or certain kinds of patients, consider the possibility of a higher cut-off, such as 0.05 or 0.1 Hz maximum."
Besides a very good skin preparation to be done before Ag-AgCl electrodes attaching, particular care must be given to 50/60 Hz electric interferences (i.e, notch filtering to reduce as much possible any residual power supply and/or environment negative interference), electrode movement artifcats (it must be verified, a posteriori, if the conducting gel has been used properly, if the skin has been suitably prepared, and so on), ECG/EMG interference reduction (if necessary, by means of an ICA - Independent Component Analysis - filtering or, alternatively, some adaptive filtering - either Wiener or Kallman or other recently published algorithms which use some a-priori knowledge of the interfering noise statistical characteristic). Very deep care must be given also to the amplifier properties of the ASC (Analog Signal Conditioning) block, whose front-end stage could be better implemented in a double-differential configuration with OpAmp with elevate CMRR (Common Mode Rejection Rate), i.e. above 140dB.
Hi Petteri (sorry for the late reply),
0.01Hz cutoff seems reasonable and so is 0.1Hz. The optimal cutoff settings depend on the thing you want to examine. If you look at very late components (e.g. P3) you should use a low value (< 0.5 Hz). It is better not to use a high-pass filter at all if you look for delay period activity in working memory tasks (e.g. CDA / SPCN). However, if you are interested in early ERP components (at ~100 ms, for example the auditory N1) you can use much higher cutoff values.
http://www.ncbi.nlm.nih.gov/pubmed/24095710
In fact, high cutoffs are sometimes used to baseline data (if conditions differ in the unfiltered data during the baseline)
The take-home message is: with a 0.1 Hz cutoff you are generally on the safe side.
cheers
Tobi
You might use ICA method for removing some artifacts such as, EMG, ECG and EOG signal. In addition, you adaptive filter can be used to reducing ECG and EOG artifacts.
You can filter your signal in the frequency domain, according to your purpose.
I think use of empirical mode decomposition-detrended fluctuation analysis is critical to remove the EOG artifact noise.
- Badges
- Science topic
- Similar topics
- Biological Science
- Physiology
- Biosignals
- Biomedical Signal Processing