Login for members

Signal to Noise in EEG

What is the signal-to-noise ratio and why should we care?

Electroencephalography (EEG) is one of the key technologies used to observe brain activity in neuromarketing research. Compared with other methods, it is precise: measuring changes in brain states that occur in thousands of a second, literally “at the speed of thought”. It is also relatively low cost and easy to use. However, using EEG is not completely without certain impediments. The signal-to-noise ratio (SNR) is one of the most important methodological challenges for EEG data collection and analysis. This article provides a short primer on SNR: where it comes from, why it matters, how it can be reduced, and what buyers need to know in order to understand how SNR is being handled in their studies

Reading time: 10 minutes


What is SNR? The best explanation is that it is the ratio of “everything you want to measure in your analysis” to “everything else picked up by the EEG signal.” This noise is a problem because there are two major sources of noise in EEG signals. The first is the general background noise that comes from outside the brain. The second is the natural noise that originates inside our brains due to the fact that our brains are always busy doing lots of things at once, not just the single aspect you want to focus on in your study. All these activities contaminate the EEG signal.

The sources of contaminating noise

Let us examine external noise first. One of the challenges of EEG technology is that electrical activity generated by the brain is miniscule, on the order of millionth of a volt. Consequently, scalp recorded electrical activity consists of a mix of genuine brain signals combined with lots of noise - termed artifact - generated by other parts of the body, such as heart activity, eye movements and blinks, other facial muscle movements, etc., which produce electrical signals about 100 times greater than those produced by the brain. So an initial task of any EEG data analysis is artifact removal, which consists of separating these other signals from the signals being emitted by the brain itself.

Another source of external noise is the environment within which the EEG data is recorded. The most common sources of this environmental noise are the ambient electrical current in any room with electrical wiring, either 50 Hz or 60 Hz (according to which country you are in, and any other electrical equipment in close proximity to the EEG sensors. These signals are usually removed from the EEG recording using a notch filter that eliminates signals at specific frequencies.

Internal noise is trickier. This is due to the fact that our brains are engaged in many different activities at any moment in time, and each of these activities generates electrical activity that gets mixed into the overall signal picked up by EEG sensors on the scalp. The problem is complicated by the fact that activity always occurs throughout the brain, both on the cortical surface near the scalp and in structures deep inside the brain, and reaches the scalp in many different ways. Disentangling all these signals to focus on a particular signal of interest for a given study is a major challenge that is most often addressed by the principles of controlled repetition and averaging.

Why does it matter?

If the signal in an EEG analysis is not properly separated from the noise surrounding it - both external and internal noise - the results are likely to be incorrect and highly misleading. It may seem that a particular response is occurring, but if you have not explicitly corrected the signal-to-noise ratio, that response is basically meaningless.

Improving the signal-to-noise ratio

Measures to control the SNR consist of two types: eliminating external noise sources and separating internal noise from the signal of interest. If possible, the best way to cope with external noise is to avoid it in the first place. To eliminate equipment noise, it is best to use high quality devices and electrodes (for instance, most dry electrodes are useless because they are very sensitive to external noise) and remove any sources of electromagnetic noise, such as cables, cell phones, laptops, computer monitors, etc. from the recording area. This is relatively easy compared to eliminating noise that is generated by the subject himself or herself. Asking the participant to sit still is a common practice, but it is essentially impossible to prevent all blinking or facial muscle movement.

This problem is often addressed by a smart experimental protocol, which can minimize internal cerebral noise by keeping the participant focused and engaged in the task at hand, while also providing frequent breaks so participants can squirm and blink periodically in between experimental tasks. Well-trained EEG technicians can also help by making the participant feel comfortable and creating a relaxed, professional atmosphere for the data collection.

After the data has been recorded, in what is called post-processing, advanced statistical algorithms are often used to identify and remove subject-related noise such as movement artifacts, eye blinks and muscle tension from the raw EEG signal. For example, machine learning algorithms can identify patterns in the signal associated with external or internal noise and separate those signals from the brain signals of interest. A family of statistical techniques called Blind Signal Separation (BSS) algorithms are often used to perform this task, and are commonly built directly into EEG analysis software, both commercial and open source. Often these techniques are accompanied by manual cleaning, during which an EEG expert examines the signal visually and removes artifacts and noisy sections by hand.

ERP: Averaging the results

The main technique used to deal with internal noise is repetition and averaging. This can best be illustrated using a subcategory of EEG analysis called Event Related Potential (ERP) analysis, a very common method used in many neuromarketing EEG studies. ERP is a signal within the EEG which reveals how information is being processed in the brain. This signal can be obtained by time-locking the recording of the EEG with the onset of an event, say the presentation of a word or image.

Repetition and averaging are used in ERP studies to separate the signal of interest from the noise of other brain activity. The underlying assumption is that the signal triggered by a stimulus event will remain relatively constant over multiple trials, while all the other signals will occur randomly across trials. Therefore, if you average multiple trials of one or more people being exposed to the same stimulus event over and over again, the signal will continue to stand out because it is not randomly distributed, but the noise will tend to average to zero, thus in effect disappearing from the average and allowing the signal you want to study to stand out in all its glory.

If you are confident that all the people in your sample are similar, you can average their individual averages together to produce what is called the Grand Average. This is what is usually reported in academic articles that use the ERP technique.

SNR: What a neuromarketing buyer should know

Here are some questions clients can and should ask to better understand how SNR issues are addressed by their neuromarketing vendor.

1. What procedures are being used for artifact correction and “data cleaning” in my study? Many procedures are available, and your vendor should be able to describe a detailed “preprocessing chain” used to find and remove noise artifacts in their raw EEG data.

2. What procedures are being used to assess the statistical significance of the obtained results? It is not sufficient simply to present the results, it is important to state the “distance” between the observed results and results that might be produced by random fluctuations due to the random noise and chance alone.

3. Can your procedures be corroborated by reference to published papers? Be suspicious if any vendor claims to have “proprietary algorithms” not based on published research. In a field built upon decades of research by thousands of researchers in thousands of labs, it is highly unlikely that a commercial vendor has discovered something brand new that has eluded everyone else in the field.

Terms and Abbreviations used

EEG = electroencephalogram, the on-going electrical activity of the brain measured from electrodes placed on the scalp

ERP = event related potential; a signal within the EEG which reveals more information about information processing in the brain. This signal can be obtained by time-locking the recording of the EEG to the onset of events

SNR = signal-to-noise ratio; the ratio between “what you would like to measure” and “all the other things”

Artifact correction = the partially automated and partially manual process of separating brain signals from other signals that can show up in the electrical recording

Machine learning algorithms = a branch of artificial intelligence that studies the construction and study of systems that can learn from data; essential tools for classification.


This article has briefly explained the signal-to-noise issue in EEG, without too many technical details. If you would like to learn more, consult the resources below:

Bang J W, Choi J S, Park K R (2013) Noise Reduction in Brainwaves by Using Both EEG Signals and Frontal Viewing Camera Images. Sensors 13:6272-6294.

Croft R J, Barry R J (2000) Removal of ocular artifact from the EEG: a review. Neurophysiologie clinique 30(1):5-19.

Delorme A, Makeig S (2004) EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics. Journal of Neuroscience Methods 134:9-21.

Delorme A, Sejnowski T, Makeig S (2007) Improved rejection of artifacts from EEG data using highorder statistics and independent component analysis. Neuroimage 34(4): 1443-1449.

Dien J (2010) The ERP PCA Toolkit: an open source program for advanced statistical analysis of eventrelated potential data. Journal of neuroscience methods 187(1):138-145.

Goldenholz D M, Ahlfors S P Hämäläinen M A, Sharon D, Ishitobi M, Vaina L M, Stufflebeam S M (2009) Mapping the Signal-To-Noise-Ratios of Cortical Sources in Magnetoencephalography and Electroencephalography. Human Brain Mapping 30:1077–1086.

Krishnaveni V, Jayaraman S, Gunasekaran A, Ramadoss K (2006) Automatic Removal of Ocular Artifacts using JADE Algorithm and Neural Network. International Journal of Biological and Life Sciences 2(1):10-21.

Li Y, Ma Z, Lu W, Li Y (2006) Automatic removal of the eye blink artifact from EEG using an ICA-based template matching approach. Physiological measurement 27(4):425-36.

Luck S (2005) An Introduction to the Event-Related Potential Technique. Cambridge MIT Press.

Olkkonen H, Pesola P, Olkkonen J T, Valjakka A, Tuomisto L (2002) EEG noise cancellation by a subspace method based on wavelet decomposition. Medical Science Monitor 8(11):MT199-204.

Repovš G (2010) Dealing with Noise in EEG Recording and Data Analysis. Informatica Medica Slovenica 15(1):18-25.

This article (textual and pictorial content) is subject to copyright. For educational purposes only. More information: Neuromarketing Science & Business Association.