noisy-signal

How To Reduce Noise In EEG Recordings [11 Solutions]

If you are working with bio-data, you have probably come across noise. But what exactly is noise and how can we remove it?

Noise is anything a sensor detects, which the researcher did not intend it to detect. All data-sets can be broken down into: 

Raw Data = Signal + Noise

Most researchers aim to maximize the Signal-to-Noise Ratio (SNR). That is, reduce noise, and increase the power of their analysis, without relying on enormous sample sizes.

Sources of noise: Where does signal noise come from?

For EEG research, external noise and artifacts are signals that do not come from the brain, but which the sensors detect. There are different sources of noise and artifacts in EEG data.

Interested in EEG?

Physiological noise

Physiological factors are known to introduce noise into EEG recordings. These include, but are not limited to:

  • Cardiac signal (Electrocardiogram, ECG or EKG)
  • Artifacts caused by muscle contraction (Electromyogram, EMG)
  • Ocular signal caused by eyeball movement (Electrooculogram, EOG)
  • Irrelevant underlying brain activity not pertaining to the experiment
Environmental noise 

Anything that uses electricity will emit an electromagnetic field that may be detected by your measuring equipment. Some examples of such external noise sources are: 

  • AC power lines
  • Room lighting and computer equipment
  • Any electronic equipment in the vicinity of the sensors
Motion artifacts

Whenever a physical part of the measurement setup is moved, it can cause visible artifacts in the data. These movements can happen at the intersection of electrodes, or when moving electrode cables.

Examples of artifacts in EEG data
4 types of noise in EEG data
Figure 1: Visual patterns of different artifacts.

1 EOG blinking artifact. Big amplitude, slow, positive wave prominent in frontal electrodes.

2 Electrode artifact caused by unstable contact (bigger impedance) between P3 electrode and skin.

3 Swallowing artifact.

4 Common reference electrode artifact caused by unstable contact between the reference electrode and skin. Huge wave similar in all channels.

Why cleaning your EEG signal is critical

Statistical analysis becomes more powerful the less noise your signal has. In other words, the cleaner the data, the more representative the analysis. Maximizing the signal-to-noise ratio (SNR) is critical to clear results.

How to increase your signal-to-noise ratio

Increasing sample size and averaging

Meisler et al. (2019) concluded that increasing your sample size instead of investing large amounts of time into data cleaning can be worthwhile. The idea: noise will average out; signals prevail.

If your recording sessions are short and participants are readily available, this approach is feasible. When your recording sessions take hours or participants are difficult to recruit, data cleaning may be a better option.

If you cannot easily increase your sample size, however, there are ways to improve your SNR without resorting to data cleaning.

Before EEG recording

There are several simple and effective measures you can take long before you start recording. If you know the factors that cause noise and artifacts, you can design your experiment in a way that minimizes these factors: 

  • If you are investigating stationary EEG, consider performing your experiment in an electromagnetically isolated room. Use a Faraday cage if your institution has one. 
  • Remove or replace any electronics that use AC (alternating current) with equipment using a DC. 
  • Ensure your participants are in a comfortable resting position to reduce ECG noise.
  • Eliminate EMG artifacts by removing tasks that require verbal responses or large movements.
Reduce ECG artifacts by allowing your participant to sit in a comfortable resting position.
  • Shorten your recording sessions. Wet electrodes tend to lose conductivity once their conductive gel starts to dry or evaporate. The rate at which this happens depends on the quality of the gel and the duration of the sessions. For long sessions, consider using dry electrodes for increased signal stability. More details about the benefits of wet and dry electrodes can be found in this article
  • Explore new experimental designs. New hardware is being developed with the express aim of increasing SNR. Nordin et al. (2018) designed a dual layer setup, in which a secondary sensor array detects motion artifacts, which can be subtracted from the main EEG data (see also Richer et al., 2020; Nordin et al., 2019; Nordin et al., 2020).
During EEG recording

If your motion artifacts are mostly caused by electrodes and cables changing position while participants move, consider minimizing moving parts:

  • Minimize cable length. Each centimetre of cable that connects the electrodes to the amplifier may introduce motion artifacts. The less cable you use, the better. At Mentalab we tailor our cables to the size of the EEG cap being used, and offer custom cable lengths.
  • Reduce cable movement. You may be able to attach any moving cables to the EEG cap using velcro, putty, or similar. Ideally, your cables will not move when the participant moves. Mentalab’s neoprene caps, for example, allow you to fix your cables with a simple and effective cap design.
  • Verify your impedances before you start measuring. Electrode impedances indicate how good the electrical contact is between the participant’s skin and the electrodes. The lower the impedance value, the more signal the electrode can detect. Mentalab’s in-house API explorepy allows you to measure impedances for each electrode of your Explore Device quickly.
Clip cables
Shorten the length of your cables to minimize noise. The shorter the cable the better.

After EEG recording

Having taken all necessary precautions before and during recording, and with a few test recordings under your belt, you may be baffled to find that there is still a lot of noise in your data.

Rest assured, EEG is tricky and this is very common. Fortunately, there are now many mathematical tools to help you increase your SNR after recording.

Let’s take a look at how you can find the signal in the noise after your recording is complete.

Manual inspection

If you have worked with EEG in the past, you probably have a good idea of what a clean signal looks like. You may even be familiar with the common artifacts represented in Figure 1.

As such, the first step to improving your SNR is to look at the plotted data. See if you can easily differentiate a clean signal from artifacts and noise. This will indicate the quality of your recording, so you can decide how much post-processing to do.

If you can see clear brainwaves, you may not need to do anything. However, if you see a lot of noise, you will probably need to apply some post processing techniques.

Mathematical approaches and signal decomposition

Since “taking a look at the data” is not a scientifically standardized procedure, researchers apply a variety of mathematical approaches to separate the signal from the noise. Here are the most common approaches to do this.

Artifact Subspace Reconstruction (ASR)

Several leading research groups have suggested using Artifact Subspace Reconstruction (ASR; Chang et al., 2018; Blum et al., 2019). ASR is an online, component-based method to effectively remove transient or large-amplitude artifacts.

The technique is capable of running in real-time and uses statistical anomaly detection to separate artifacts from EEG signals in multichannel data sets. It assumes that non-brain signals introduce a large amount of variance to the data set and can be detected via statistics.

ASR decomposes short segments of EEG data and contrasts them to calibration data. EEGLab, a popular (and free) analysis plugin for Matlab developed by UCSD, offers ASR as a feature.

Independent Component Analysis (ICA)   

Wu et al. (2018) define Independent Component Analysis (ICA) as

A blind source separation technique that effectively decomposes the multichannel EEG data into multiple independent components belonging to either artifacts or neural sources, building on the observation that artifact and neural signals possess distinguishable spatio-temporal patterns. 

In other words: if you visualize signal and noise, the noise looks very different. These differences can be framed and used to categorize components as either signal or noise. Then, noise can be removed.

For more on ICA, see Makeig et al. (1995), Makeig et al. (1997), and Makeig et al. (2002)

Canonical Correlations Analysis (CCA)

Canonical Correlations Analysis (CCA) uses autocorrelations within a given time series to characterize the signal component of an unclean data set.

Autocorrelation is the correlation between a signal and a lagged copy of itself over successive time intervals. Data points that were taken at a similar time usually have a stronger autocorrelation.

CCA looks for the most correlated components of data sets using cross-covariance matrices. In general, this technique can distinguish noise from brain signals based on relatively low autocorrelation values.

In certain scenarios, CCA can outperform ICA (Gao et al., 2010), and there are techniques to use CCA in real-time (Lin et al., 2018). Most other techniques can only be applied after a recording is finished.

For event-related potentials research, you can use CCA to create linear models of a stimulus and its brain response (de Cheveigné et al., 2018).

Multiple Sparse Priors (MSP)

If you are working with event-related potentials, you may also want to carry out source-localization. A relatively new and promising approach to accurately locating EEG signals is to use Multiple Sparse Priors (MSP).

MSP uses Bayesian statistics and has outperformed conventional methods in some studies (for a detailed breakdown see; Friston et al., 2008).

For more on MSP, and how it compares to other methods, have a look at our article on source localization and the event-related potentials inverse problem.

Auto-rejection approaches

When working with higher density EEG systems, you may find that individual electrodes are not recording a clean signal, which can affect the overall quality of the data set.

To solve this problem, there are different approaches to automatically detect and reject “bad” channels in real-time. To avoid incomplete data arrays after removing individual channels, researchers tend to interpolate bad channels’ signals using sets of neighboring channels that have been identified as “good”.

Sensor Noise Suppression (SNS)

Sensor Noise Suppression (SNS) assumes that true signals will be picked up by more than one sensor. As such, each channel is projected onto the subspace spanned by its neighbors and replaced by its projection.

That is, SNS removes a channel’s noise by seeing what bits of its signal are unique and what bits show up in other channels. Noise will not show up, signals will. Again, the idea here is that brain signals will project onto multiple sensors, while noise is uncorrelated across sensors.

For high density systems, not all channels are used for projection, but only a subset of neighboring channels (determined by the correlation between them). To learn more about this technique see de Cheveigné and Simon (2008).

Fully Automated Statistical Thresholding for EEG artifact Rejection (FASTER)

Fully Automated Statistical Thresholding for EEG Artifact Rejection (FASTER) uses five different statistical criteria to identify “bad” sensors. These criteria are: variance, correlation, the Hurst exponent, kurtosis and line noise.

A sensor is bad in one criterion if its z-score exceeds 3 in that criterion. See Nolan et al. (2010) for more.

Random Sample Consensus (RANSAC)

Random Sample Consensus (RANSAC) is a well-known approach used to fit statistical models to data that contain outliers. The method was first described by Fischler and Bolles (1981). However, it was Bigdely-Shamlo et al. (2015) who adopted RANSAC to create a standardized, early-stage EEG processing pipeline (PREP).

In their approach, a random subset of sensors (25% of the sensors, called “inliers”) are sampled. The data in all sensors are interpolated from these inlier sensors. This is repeated multiple times (50 in the PREP implementation) resulting in sets of 50 time series for each sensor. 

The correlation between the medians of these 50 time series and the real data is computed. If the correlation value falls below a set threshold (0.75 in the PREP implementation), the sensor is considered an outlier and labeled “bad”.

If a sensor is “bad” for over 40% of trials, it is marked as globally bad and interpolated. The PREP pipeline can be downloaded as a free Matlab/EEGLab library here.

Cross-validation

Jas et al. (2017) use cross-validation on signal data. They dynamically adapt peak-to-peak thresholds based on the characteristics of the data to detect and reject bad channels. They then replace bad data using interpolation from neighboring channels where possible.

When used to repair bad segments in four different open data sets containing over 200 subjects, this technique performed as well as or better than the conventional auto-rejection approaches mentioned above.

Deep learning

Finally, deep learning uses existing data sets to train an AI algorithm to automatically detect and reject artifacts from a recording. Ideally, this is done in real-time.

For instance, Mashhadi et al. (2020) successfully removed noise from their signal reliably and accurately using deep learning. They aimed to reduce the mean square error between the target signal (pure EEG) and the predicted signal (purified EEG).

Yang et al (2019) went further and created a human workload analyzer using EEG and deep learning. They introduced a feature that found characteristics of high dimensional EEG indicators. They concluded that, once the system had identified an optimal network architecture, it outperformed several other methods.

Conclusion

All these techniques have one goal in common: increasing SNR in the most efficient way. They all deliver on their objective, but some are more suitable for certain scenarios than others.

Since many analysis platforms now offer some or several of these methods, best practices will emerge in the near future. There is also a trend towards real-time data analysis, driven by new software that saves researchers time and effort.

Hardware components are also being modified to improve researchers’ SNR. These include the development of shielded cables and active electrodes, as well systemic approaches like dual layer setups.

However, the “holy grail” of clean EEG signal acquisition, especially for mobile solutions, will be sensors that don’t require any cables at all. “Never say never” might hold here, but we are not sure if and when wireless mobile EEG will become a reality.

We are curious to hear about your experiences with different EEG data processing. Please reach out to us at

References

Blum, S., Jacobsen, N., Bleichner, M. G., & Debener, S. (2019). A Riemannian Modification of Artifact Subspace Reconstruction for EEG Artifact Handling. Frontiers in human neuroscience, 13, 141. https://doi.org/10.3389/fnhum.2019.00141

Bigdely-Shamlo, N., Mullen, T., Kothe, C., Su, K. M., & Robbins, K. A. (2015). The PREP pipeline: standardized preprocessing for large-scale EEG analysis. Frontiers in neuroinformatics, 9, 16. https://doi.org/10.3389/fninf.2015.00016

Chang, C. Y., Hsu, S. H., Pion-Tonachini, L., & Jung, T. P. (2018). Evaluation of Artifact Subspace Reconstruction for Automatic EEG Artifact Removal. Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Annual International Conference, 2018, 1242–1245. https://doi.org/10.1109/EMBC.2018.8512547

de Cheveigné, A., & Simon, J. Z. (2008). Sensor noise suppression. Journal of neuroscience methods, 168(1), 195–202. https://doi.org/10.1016/j.jneumeth.2007.09.012

de Cheveigné, A., Wong, D., Di Liberto, G. M., Hjortkjær, J., Slaney, M., & Lalor, E. (2018). Decoding the auditory brain with canonical component analysis. NeuroImage, 172, 206–216. https://doi.org/10.1016/j.neuroimage.2018.01.033

Fischler, M.A., & Bolles, R.C. (1981). Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24, 381-395. https://doi.org/10.1145/358669.358692

Friston, K., Harrison, L., Daunizeau, J., Kiebel, S., Phillips, C., Trujillo-Barreto, N., Henson, R., Flandin, G., & Mattout, J. (2008). Multiple sparse priors for the M/EEG inverse problem. NeuroImage, 39(3), 1104–1120. https://doi.org/10.1016/j.neuroimage.2007.09.048

Gao, J., Zheng, C., & Wang, P. (2010). Online removal of muscle artifact from electroencephalogram signals based on canonical correlation analysis. Clinical EEG and neuroscience, 41(1), 53–59. https://doi.org/10.1177/155005941004100111

Lin, C. T., Huang, C. S., Yang, W. Y., Singh, A. K., Chuang, C. H., & Wang, Y. K. (2018). Real-Time EEG Signal Enhancement Using Canonical Correlation Analysis and Gaussian Mixture Clustering. Journal of healthcare engineering, 2018, 5081258. https://doi.org/10.1155/2018/5081258

Jas, M., Engemann, D. A., Bekhti, Y., Raimondo, F., & Gramfort, A. (2017). Autoreject: Automated artifact rejection for MEG and EEG data. NeuroImage, 159, 417–429. https://doi.org/10.1016/j.neuroimage.2017.06.030

Makeig, S., Bell, A. J., Jung, T.-P., & Sejnowski., T. J. (1995). Independent component analysis of electroencephalographic data. Proceedings of the 8th International Conference on Neural Information Processing Systems (NIPS’95), 145–151.

Makeig, S., Jung, T. P., Bell, A. J., Ghahremani, D., & Sejnowski, T. J. (1997). Blind separation of auditory event-related brain responses into independent components. Proceedings of the National Academy of Sciences of the United States of America, 94(20), 10979–10984. https://doi.org/10.1073/pnas.94.20.10979

Makeig, S., Westerfield, M., Jung, T. P., Enghoff, S., Townsend, J., Courchesne, E., & Sejnowski, T. J. (2002). Dynamic brain sources of visual evoked responses. Science (New York, N.Y.), 295(5555), 690–694. https://doi.org/10.1126/science.1066168

Mashhadi, N., Khuzani, A.Z., Heidari, M., & Khaledyan, D. (2020). Deep learning denoising for EOG artifacts removal from EEG signals. 2020 IEEE Global Humanitarian Technology Conference (GHTC), 1-6. https://doi.org/10.1109/GHTC46280.2020.9342884

Meisler, S. L., Kahana, M. J., & Ezzyat, Y. (2019). Does data cleaning improve brain state classification?. Journal of neuroscience methods, 328, 108421. https://doi.org/10.1016/j.jneumeth.2019.108421

Nolan, H., Whelan, R., & Reilly, R. B. (2010). FASTER: Fully Automated Statistical Thresholding for EEG artifact Rejection. Journal of neuroscience methods, 192(1), 152–162. https://doi.org/10.1016/j.jneumeth.2010.07.015

Nordin, A. D., Hairston, W. D., & Ferris, D. P. (2018). Dual-electrode motion artifact cancellation for mobile electroencephalography. Journal of neural engineering, 15(5), 056024. https://doi.org/10.1088/1741-2552/aad7d7

Nordin, A.D., Hairston, W.D., & Ferris, D. P. (2019). Human electrocortical dynamics while stepping over obstacles. Sci Rep 9, 4693. https://doi.org/10.1038/s41598-019-41131-2

Nordin, A. D., Hairston, W. D., & Ferris, D. P. (2020). Faster Gait Speeds Reduce Alpha and Beta EEG Spectral Power From Human Sensorimotor Cortex. IEEE transactions on bio-medical engineering, 67(3), 842–853. https://doi.org/10.1109/TBME.2019.2921766

Richer, N., Downey, R. J., Hairston, W. D., Ferris, D. P., & Nordin, A. D. (2020). Motion and Muscle Artifact Removal Validation Using an Electrical Head Phantom, Robotic Motion Platform, and Dual Layer Mobile EEG. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 28(8), 1825–1835. https://doi.org/10.1109/TNSRE.2020.3000971

Wu, W., Keller, C. J., Rogasch, N. C., Longwell, P., Shpigel, E., Rolle, C. E., & Etkin, A. (2018). ARTIST: A fully automated artifact rejection algorithm for single-pulse TMS-EEG data. Human brain mapping, 39(4), 1607–1625. https://doi.org/10.1002/hbm.23938

Yang, S., Yin, Z., Wang, Y., Zhang, W., Wang, Y., & Zhang, J. (2019). Assessing cognitive mental workload via EEG signals and an ensemble deep learning classifier based on denoising autoencoders. Computers in biology and medicine, 109, 159-170. https://doi.org/10.1016/j.compbiomed.2019.04.034