I have EEG recordings of an experiment. In some of these recordings subject made an error during the experiment. We know these error cause brain to output ErrP signals. I want to detect this. There are not many sources about single trial detection of ErrP's. What are the main steps to detect these error indicator signals?

Error-related potentials (ErrPs) are neurophysiological signals associated with error processing. They are generated when wrong actions are perceived and have been reported in many contexts in the past two decades, namely when a subject perceives that he/she has committed an error and recognizes it immediately (‘response ErrP’), when a subject receives the feedback of a previous choice without knowing whether it was wrong (‘feedback ErrP’), when observing mistakes of another person or agent (‘observation ErrP’)’ or during the interaction with a brain-computer interface (BCI) when the feedback is not the expected one (‘interaction ErrP’). The components of an ErrP appear within a time window of 500 ms and are naturally elicited in the brain without the user’s explicit intention. Thus, its automatic detection can be used in myriad ways, in real-time, and in human-machine interaction processes. In particular, interaction and observation ErrPs have been applied as a proof-of-concept in several applications, for example, for detection and correction of BCI choices to increase reliability, to adapt BCI systems over the time, or to make intelligent systems (e.g., external agents) learn. There is also a growing interest for ErrPs in clinical applications for disorders where error monitoring is impaired.

  • 1
    $\begingroup$ It seems you are quoting a long passage from some source. It is imperative that you cite the source when you do this, to give proper credit and to avoid plagiarism. $\endgroup$
    – Bryan Krause
    Commented Feb 17, 2022 at 19:41

1 Answer 1


An "error related potential" is just a difference in EEG traces between error and non-error trials.

To calculate an (anything)-related potential, you average trials with and without (anything), and subtract one from the other. It's possible to do this in time or frequency domains, though you'll likely get different answers due to evoked vs induced components (see David, O., Kilner, J. M., & Friston, K. J. (2006). Mechanisms of evoked and induced responses in MEG/EEG. Neuroimage, 31(4), 1580-1591. ). If you want to isolate errors specifically, you'll want to match trials where other things are constant. For example, if your subjects are choosing left or right, you may want to subtract trials where they chose left as a correct option from those where they chose left as an error. The exact way of doing this depends entirely on your experimental paradigm. Read the literature for ideas, but the specifics depend on your specifics.

Since you know which trials are errors, you do not need to detect errors on single trials to demonstrate if some error signal is present.

If you want to do single-trial detection, that will be a matter of finding the features in the average error signal in a single trial or identifying non-phase locked features (such as power changes) that don't show up well in averages. You might look for power in particular frequencies, voltage at a specific moment in time relative to a stimulus/decision, phase relationships on different channels, etc. You could train a neural network (or use other AI/ML approaches) to detect these features using standard approaches in those fields (cross-validation, etc).

Single-trial detection with EEG is very difficult. I suspect you will fail to distinguish on single trials from EEG with high accuracy despite what you've read, but you're welcome to try, and can probably exceed chance performance if the quality of recordings and strength of signal are good.

Here are a few papers where people have done similar things:

Chavarriaga, R., & Millán, J. D. R. (2010). Learning from EEG error-related potentials in noninvasive brain-computer interfaces. IEEE transactions on neural systems and rehabilitation engineering, 18(4), 381-388.

Ferracuti, F., Casadei, V., Marcantoni, I., Iarlori, S., Burattini, L., Monteriù, A., & Porcaro, C. (2020). A functional source separation algorithm to enhance error-related potentials monitoring in noninvasive brain-computer interface. Computer Methods and Programs in Biomedicine, 191, 105419.

Salazar-Gomez, A. F., DelPreto, J., Gil, S., Guenther, F. H., & Rus, D. (2017, May). Correcting robot mistakes in real time using EEG signals. In 2017 IEEE international conference on robotics and automation (ICRA) (pp. 6570-6577). IEEE.

None of their approaches are identical; there is no "right" answer here.

  • $\begingroup$ Thanks for this awesome reply. I saw it the first time you posted it but I realized I lack general information and I read at least 20 articles about this topic. And now I can understand what you mean more clear. $\endgroup$
    – Enes Kuz
    Commented Mar 1, 2022 at 9:53
  • $\begingroup$ I also realized my interest is in single trial detection of ErrP signals. Lets say we have an epochs of EEG where t=0 is onset of an event and the epochs ends in t=500. And we know this epochs have error signals inside of them. And we have epochs with no error potentials. And we would like to create an application which will classify a random epoch as includes error signal or correct. Does it make sense to feed raw signals without any feature extraction to basic neural network? $\endgroup$
    – Enes Kuz
    Commented Mar 1, 2022 at 10:00
  • $\begingroup$ @EnesKuz Probably not, but I wouldn't say it's impossible either, given enough training data. Probably need to train separately on every subject. I think your bigger issue will be simply on the recording side: EEG is a very noisy modality. I wouldn't reinvent the wheel on this, though, there are lots of papers where people have attempted similar things, it just seems the accuracy is usually pretty crap. People will spend years and a PhD thesis on questions like this. $\endgroup$
    – Bryan Krause
    Commented Mar 1, 2022 at 15:05
  • $\begingroup$ Yeah it is not an easy task. That's why I am not concerned about accuracy score. I think doing this will be a good start for my future research. $\endgroup$
    – Enes Kuz
    Commented Mar 3, 2022 at 11:01

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.