SEREEGA: Simulating event-related EEG activity

TLDR: The architecture and general workflow of this toolbox, as well as a simulated data set demonstrating some of its functions, are presented, indicating that SEREEGA is a general-purpose toolbox to simulate ground-truth EEG data.

About: This article is published in Journal of Neuroscience Methods.The article was published on 2018-11-01 and is currently open access. It has received 35 citations till now. The article focuses on the topics: Electroencephalography.

Figures

Figure 5: ERP image (Makeig, Debener, et al., 2004) at channel Fz, Cz, and POz showing the epoch-to-epoch variability of the simulated ERP signals as well as their spatial differences and effects over time.

Figure 1: Impression of the SEREEGA GUI.

Figure 3: Left: The location of source number 7479 in the New York Head as plotted onto EEGLAB’s standard head model. Right: Four projection patterns of that source onto the scalp. From top left to bottom right: projection in the x-positive, y-positive, and z-positive direction, and a linear combination of the three (arbitrary units). This effectively orients the dipole at the source’s location in 3D space.

Figure 2: Channel locations and estimated brain boundaries for three head models. Left: The New York Head with its 228 channels. Middle: a 339-channel layout around a standard average head model available by default in FieldTrip. Right: The Pediatric Head Atlas (8 to 18, version 2) showing 2562 channels.

Table 2: Classification accuracies for two classifiers, one using the windowed means approach (WM) and one based on band power (BP). TP = true positive rate, TN = true negative rate, Acc = overall accuracy. Significance is reached at an accuracy of 0.57. The WM classifier for participant 5 does not reach significance.

Figure 4: Sample signals annotated with the base variables that control their behaviour. From top left to bottom right: event-related potential (ERP); coloured noise; single-frequency unmodulated event-related spectral perturbation (ERSP); frequency band unmodulated ERSP; inverse burst-modulated ERSP; amplitude-modulated ERSP. Each base variable can additionally be given a deviation and a slope to control the epoch-to-epoch and systematic progressive variability of the signal. Not illustrated: the autoregressive model (ARM) and data classes.

Frequently asked questions

Q1. What are the contributions in "Sereega: simulating event-related eeg activity" ?

The authors present SEREEGA, Simulating Event-Related EEG Activity. This paper presents the architecture and general workflow of this toolbox, as well as a simulated data set demonstrating some of its functions. SEREEGA unifies the majority of past simulation methods reported in the literature into one toolbox. 

Q2. What have the authors stated for future works in "Sereega: simulating event-related eeg activity" ?

In the future, the authors may consider a public database of such recordings for all users to draw from. When taking data from empirical recordings, potential anatomical differences should be taken into account. This refers to the core steps of each simulation—further implementation details may of course differ. With the consistent and increasing popularity of EEG, there is an accompanying need to further develop and validate EEG analysis methods. 

Q3. What can be done with the SEREEGA toolbox?

FieldTrip-generated lead fields based on standard or custom head models can be used, and the toolbox’s architecture allows it to be readily extended with additional head models and signals. 

Q4. What functions can be used to generate a random number of components?

convenience functions exist to generate any number of ‘random’ ERP or ERSP classes, based on a range of allowed base values. 

Q5. What is the function that outputs the simulated scalp data?

A simulation function takes the defined components, the lead field, and the general configuration as input, and outputs the simulated scalp data in a channels× samples× epochs matrix, as well as components × samples × epochs source data. 

Q6. What was the method used for the P3 classifier?

The P3 classifier followed a windowed means approach (Blankertz, Lemm, Treder, Haufe, & Müller, 2011), using the mean amplitude of six consecutive 50 ms time windows starting 200 ms after stimulus onset as features. 

Q7. What is the limitation in the current architecture?

A limitation in the current architecture is that the defined components are necessarily independent at runtime: the procedurally generated activity of one component cannot presently depend on the activity in another. 

Q8. What is the inverse burst modulation for the left motor cortex?

In the target condition, the left motor cortex classes are given an inverse burst modulation centred around 650 ms (mu; width 600 ms, taper .5, relative amplitude .5) and 600 ms (beta; width 500 ms, taper .8, relative amplitude .5) respectively. 

Q9. What was the classification of the manual response?

The manual response was classified using a logBP approach (Pfurtscheller & Neuper, 2001), using as features the power between 7 and 27 Hz, 400 to 1000 ms after stimulus onset, focused on sites covering the motor cortex: FC1, FC2, FC3, FC4, FC5, FC6, C1, C2, C3, C4, C5, and C6. 

Q10. What was the deviation of the ERSP parameters relative to the stimulus onset?

ERSP parameters had a deviation of 10% relative to stimulus onset, with the exception of latency deviation, which was fixed at 100 ms. 

Q11. What are the main issues that require careful attention when interpreting raw EEG recordings?

This and other issues including volume conduction, the placement and distance of the electrodes relative to the cortical generators of the activity they measure, and the complex relation between cortical functions and features of scalp potentials, require that great care is taken when analysing and interpreting raw EEG recordings. 

Q12. What would have to be written before testing methods that require such types of signal generators?

Additional signal classes or component properties would have to be written before testing methods that require such types of signal generators. 

Q13. How did the authors implement their methods?

The authors of these examples all implemented simulation approaches from scratch, usually by linearly mixing a number of independent signals. 

Q14. What was the amplitude of brown noise added to each component?

To simulate background processes, brown noise was added to each component (as per Freeman, Ahlfors, & Menon, 2009), with an amplitude of 5 µV. 

Q15. How is the oriented activation vector kt projected through the lead field?

the oriented activation vector ŝkt is projected through the lead field which correspondsto the source of the activation, by multiplication with the projection matrix 

Q16. What is the vector of the signal from the source to the scalp?

As + ,with x denoting the vector of the recorded or simulated scalp signal, s the source activation signal, A the projection matrix used to project signals from the source to the scalp electrodes, and denoting a vector of noise.