EEG: cleaning, ERPs, scalp maps, statistics and time–frequency

Where the positions come from (Source)

• From the files, others by name (the default): positions stored in the file come first (EEGLAB chanlocs, FieldTrip elec, BrainVision coordinates, BIDS electrodes.tsv); channels without one are placed by name on the 10-5 system.
• By name (10-5 system): every channel by its name; the file's positions are ignored.
• From a positions file: click Positions file… and choose the file of your cap or digitizer (positions file formats); its positions are matched to the channels by name and used instead of the file's.

The 10-5 template

The 345 positions of the 10-5 system^[37], which holds every 10-20 and 10-10 position, computed on an idealized spherical head with the equator through Nz, T9, Iz and T10^[38]. These are idealized directions, not the head of your participant: for source analysis, measure the positions.

Names

Names match in any case (FP1 = Fp1); the old names T3 / T4 / T5 / T6 are read as T7 / T8 / P7 / P8; a leading EEG and a reference suffix (-REF, -LE, -AR, -AVG, -A1, -A2, -M1, -M2) are dropped, so EEG Fp1-REF is Fp1. Bipolar names (Fp1-F7), A1, A2, M1, M2 (ear lobes, mastoids), EOG, ECG, EMG and other non-scalp channels get no position.

One orientation

Every position is turned to x = right ear, y = nose, z = up (the frame of BrainVision, BESA, MNE-Python and MNI) from the frame its file states (EEGLAB, FieldTrip, BrainVision, BIDS) and whatever the unit; on the scalp the direction from the centre of the head is used. So the same electrode from different files lands in the same place. Orientation check: when 3 or more channels with positions in the file also have 10-5 names, their positions are compared with the template; if they look turned by 90° (a file that does not say its axes), the other orientation is used and a note says so; if neither fits, a note asks you to check the drawing.

Rodent (skull) layouts

Positions that all lie at one height are a skull layout in mm from bregma: AP anterior +, ML right +. The drawing shows them on a skull outline with bregma and the midline, and the table has the columns AP (mm) and ML (mm) to edit. For a recording without positions (EDF, BDF, XDF, plain .mat), load a positions file: a .csv with the columns name, ap, ml.

The check (Status column) and editing

• placed; renamed (T3 → T7) (found under its 10-20 name or after cleaning the name); no position; same place as another channel (the same 10-5 name, or less than 2° apart; 0.2 mm on a skull); outside the head (more than 120° from the vertex, that is far below the ears; on a skull more than 15 mm from bregma). The notes below the table say what was assumed (axes, units) and what was turned.
• In the scalp table type a 10-5 name in As to place a channel there (for example Iz for a channel called X1); empty it to leave the channel without a position. In the skull table type AP (mm) and ML (mm). The drawing and the check update at once.
• Use this layout confirms the layout, closes the window and updates the line under the button and the Overview; Cancel keeps the layout as it was. A session keeps the source, the positions file and your edits, and the methods text says where the positions came from (only once confirmed).

Without positions

Everything except scalp maps works without positions: cleaning, trials, ERPs, measures, statistics and time–frequency use the channel names. Scalp maps use the layout shown in the line under Electrode layout… (step 1); until you confirm it with Use this layout, the maps say so. BioSemi (A1–D32) and EGI HydroCel templates are not built in yet (their licences are being checked): load the manufacturer's coordinate file as a positions file (the BioSemi Site Theta Phi list, the EGI .sfp).

A fixed order

Every button rebuilds from the files in a fixed order: bad channels → filters → reference → cut into trials → reject trials, so changing an earlier setting and clicking again never filters twice, and a session replays the same steps.

Bad channels

Bad channels are left out of the average reference, the trial rejection, the ERPs and the measures (their lines are missing in the butterfly view; the grand average of a channel uses the participants where it is good). They are not re-referenced: their data stay as recorded. Suggest flags channels whose typical spread (the median over 1-s pieces of the standard deviation) lies more than 5 robust z-scores from the other channels, or below 0.5 µV (flat); check them on the butterfly view before accepting. In EEG-BIDS files, channels with status bad in channels.tsv are marked bad.

Filters

Zero-phase FIR filters (Hamming-windowed sinc) designed following the recommendations of Widmann, Schröger and Maess (2015)^[36]. The value typed is the passband edge; the transition below a high-pass is min(max(0.25 × f, 2 Hz), f) and above a low-pass min(max(0.25 × f, 2 Hz), Nyquist − f), so the −6 dB cutoff lies in the middle of the transition, half a transition beyond (0.1 Hz high-pass: −6 dB at 0.05 Hz; 30 Hz low-pass: −6 dB at 33.75 Hz). The order of each filter is 3.3 × sampling rate / transition, rounded up to an even number (0.1 Hz at 500 Hz: 16 501 taps, 33 s); a band-pass is the high-pass and the low-pass combined into one filter. The filter delays nothing (its taps are symmetric). So filter the continuous recording, not short trials; at each end the recording is extended by its mirror image turned upside down (so level and slope continue smoothly) for half the filter length. A high-pass of 0.1 Hz keeps slow ERP components such as the P300; higher values (1 Hz) distort them. The notch removes 50 or 60 Hz and its harmonics, each from 0.5 Hz below to 0.5 Hz above (wide enough for the small drift of the mains frequency), with 1 Hz transitions (−6 dB at ± 1 Hz, everything outside ± 1.5 Hz kept); a low-pass below the line frequency already does. The methods text gives the edges, transitions, −6 dB cutoffs and order of every filter.

Reference

Average subtracts the mean of the good channels from every good channel; Linked mastoids the mean of TP9 and TP10 (or M1 / M2, A1 / A2); Channels the mean of the channels typed (a single channel becomes 0). The online reference (for example FCz) is not in the data and cannot come back. The Overview and the methods text give the reference used.

ICA, interpolation and advanced cleaning

Eye movements, muscle and bridged channels are not cleaned here: do it in EEGLAB (pop_runica, ICLabel, pop_interp) or FieldTrip (ft_componentanalysis, ft_channelrepair) and load the cleaned file (.set or FieldTrip .mat); its history is read into the Overview and the methods text. Load it as it is and skip step 2, or use step 3 only to reject trials.

Trials

• Events and names: event types match ignoring case and repeated spaces (S 1 finds the BrainVision marker S  1); the name given becomes the trial's condition. Trials across a gap in the recording (EEGLAB boundary events) are left out, as are events too close to the start or end (the Overview says how many).
• Reject trials: a trial is rejected when, on any good channel, its peak-to-peak amplitude (largest minus smallest value in the trial) or its absolute amplitude (largest distance from 0 µV) exceeds the threshold. 100 µV peak-to-peak after a 0.1–30 Hz band-pass is a common start for adults; if many trials go, look at the channels named in the info line (mark them bad) rather than raising the threshold. The methods text gives the event names and the trials rejected per condition.

• Trials per condition: the fewest trials left in any condition: Check below 20 (enough for a large component such as the P300, noisy for small ones such as the N1 or an N400 difference), Warning below 10 (an ERP from fewer than 10 trials is mostly noise).
• Rejection balance (when the trials were rejected here): the share of each condition the rejection left out: Check when two conditions differ by more than 20 percentage points (and the more rejected one lost at least 3 trials), Warning above 40 points. Artefacts that come with one condition (for example blinks after targets) leave trials that are no longer comparable. A Note when the trials were rejected before loading (the history says so): the share of each condition is then not known.
• Condition balance: one condition with more than 2 times as many trials as another is a Check with the peak amplitude (the average of fewer trials keeps more noise, so its peaks come out larger) and OK with the mean amplitude, which the number of trials does not bias (rare conditions are the point of an oddball design).
• Bad channels: Check above 10% of the channels, Warning above 20%. Bad channels are left out of the ERPs, the measures and the average reference, so the reference and the scalp maps rest on fewer channels; so many bad channels often mean a poor recording (cap fit, gel, impedances).
• Interpolated channels (read from the history: EEGLAB pop_interp, FieldTrip ft_channelrepair): Warning when every measured channel (step 5, else the channels of step 4) was interpolated in a participant: the value is an estimate from the neighbours, not a recording (peaks are smoothed and the neighbours' signals count twice); Check when some of them were, or when the history does not say which channels; OK otherwise.
• Checks of the test (after Compare conditions, step 6; listed after the rows above, gone with a new measure): the same checks as Signal Characterization's statistics, for participants. Sample size: Check below 8 participants (one participant can change the verdict), Warning below 3 or when the rank-based test reported cannot give p below 0.05 with this n (exact Wilcoxon: 2 / 2ⁿ, e.g. 0.0625 for 5 participants). Normality: Shapiro–Wilk on the paired differences (2 conditions) or on the residuals (3 or more: each value minus its participant's and its condition's mean): Check at p < 0.05, Warning when the rank-based test also gives another verdict; OK with a rank-based result; the most extreme participant is named. Sphericity (repeated-measures ANOVA, 3+ conditions): Check when Mauchly's test rejects it (the Greenhouse–Geisser corrected p is reported), Warning when the correction changes the verdict at 5%. Robustness check: a Check when the parametric and the rank-based tests disagree. Missing values: a Check when participants were left out because they have no value in a condition (no trials left in it). The Statistics tab sums them up.

OK is green, Check amber, Warning red, Note grey. The checks are saved in the session, listed on the PDF report (warnings first) and summed up in the methods text.

Colours

All maps share one colour scale, symmetric about 0 µV: red positive, blue negative, white 0, with contour lines.

Scalp and skull layouts

A scalp layout gives a map of the head seen from above (nose up), out to the outermost electrode; values outside the electrodes are extrapolated and less certain. A skull layout (mm from bregma) gives a flat map drawn only inside the outline of the electrodes, because nothing is known outside them: a few skull screws give a coarse map, so read it as a smooth guess between the screws. How the values between the electrodes are computed is under Methods.

Electrodes used and left out

Electrodes used are black dots; channels left out (bad channels, channels without a position or without a value) are grey rings or are not drawn, and the line above the maps lists them. At least 3 electrodes are needed (on a skull, not all on one line).

The reference

Maps show voltages against the reference of the data: an average reference makes the map sum to about zero over the electrodes, so a positive focus comes with negative values elsewhere; a mastoid reference gives other maps of the same data.

Wavelet cycles

Wavelet cycles set the trade-off: a wavelet of n cycles at f Hz lasts about n / f s (3 cycles at 10 Hz: 0.3 s, cut at ±0.14 s). Fewer cycles give finer timing and coarser frequencies; more cycles the opposite. 3 cycles suit trials of about 1 s; 5 to 7 cycles longer trials.

Grey means no value

A value is kept only where the whole wavelet lies inside the trial, so it is exactly what a longer recording would give (nothing is padded or mirrored). Low frequencies have long wavelets, so they lose more of the trial edges, and a frequency whose wavelet never fits inside the Baseline has no ERSP. The line under the button gives the lowest frequency with values and with a baseline. For time–frequency, cut longer trials (step 3, for example −600 to 1000 ms; trials may overlap) or export longer trials from EEGLAB / FieldTrip.

Bands

Band power uses the frequencies of the Band: Delta 1–4, Theta 4–8, Alpha 8–13, Beta 13–30, Gamma 30–80 Hz. These limits are conventions that differ between labs and species. The baseline is shaded; the shade around each line is the SEM across trials (one participant) or across participants (grand average).

ERPs

• The ERP of a condition is the mean of its trials, per channel. With Subtract a baseline, the mean of the baseline window is first subtracted from every trial and channel.
• The grand average is the mean of the participants' ERPs, so every participant counts once whatever their number of trials. Its shade is the SEM across participants; for one participant the shade is the SEM across trials.
• Several channels typed together are averaged into one waveform (a region of interest).

Measures

• Mean amplitude: the average voltage in the window. It is robust to noise and does not depend on the number of trials, so it is the usual choice.
• Peak amplitude: the largest positive or negative value in the window and its latency. Peaks grow with noise, so compare them only between conditions with similar numbers of trials.
• Choose the window before looking at condition differences, from the grand average of all conditions or from the literature. Choosing it where the conditions differ most, and then testing that difference, inflates the effect ("double dipping").

Scalp maps

• Scalp layouts: spherical splines^[39], order m = 4 with 50 Legendre terms and no smoothing, so the map passes exactly through every electrode's value. Perrin et al. write the map as a constant plus one term per electrode and solve for them so that the map equals each electrode's value. With 7 terms instead of 50 the map changes by at most a few percent, beyond the outer electrodes.
• Skull layouts (mm from bregma): a thin-plate spline in the plane^[40] with a linear term, drawn only inside the outline of the electrodes, because nothing is known outside them.
• The maps are computed in core/ScalpMap.m.

Statistics

• One value per participant and condition, compared within participants. Three or more conditions: repeated-measures ANOVA with Mauchly's test and the Greenhouse–Geisser correction when needed, then Holm-corrected paired t-tests. The other test family runs as a robustness check. The computations are in core/EEGAnalysis.m and core/GroupStats.m (base MATLAB, no toolboxes).

Time–frequency

• Each trial is convolved with complex Morlet wavelets (a sine wave inside a Gaussian; the same as in LFP Analysis), one per frequency in 1 Hz steps, normalised to unit energy and cut at ±3 standard deviations of the Gaussian. The mean of each trial is removed first. The power (squared magnitude) is averaged over the trials of each condition and over the Channels (empty = those of step 4; bad channels are left out).
• ERSP (event-related spectral perturbation)^[41]: 10 × log10 of the power over its mean in the Baseline, per condition and frequency, in dB (+3 dB: power doubled; −3 dB: halved; −6 dB: a quarter). It shows power changes whether or not they are phase-locked to the event. A minus B is the difference of the two ERSPs, in dB.
• Phase locking (ITPC, inter-trial phase coherence)^[42]: the length of the mean of the trials' unit phase vectors, from 0 (random phases) to 1 (the same phase in every trial). Evoked potentials are phase-locked; an alpha decrease is not. By chance ITPC is about 0.9 / √trials (0.14 for 40 trials, 0.23 for 15), so compare conditions with similar numbers of trials; the line above the images gives the chance level of each.
• Band power: the mean wavelet power over the frequencies of the Band, as % change of each trial from the condition's mean band power in the baseline^[43]. Values only where every frequency of the band fits.
• The grand average is the mean of the participants' ERSP (in dB), ITPC and band power, every participant once. The computations are in core/EEGAnalysis.m and core/TimeFrequency.m.

• Data (Try demo data): an oddball study, 8 participants, 32 channels at 250 Hz, trials from −200 to 800 ms, already cleaned (band-pass 0.1–30 Hz, average reference, ICA components 1 and 3 removed, 5 trials rejected, T7 interpolated: the Overview tab lists these steps). Conditions Standard (40 trials), Target (15) and Novel (15) before rejection.
• Electrode layout (step 1): the positions come from the file (a spherical head): 32 of 32 channels placed: 32 from the file. Every Status is placed; none is renamed or without a position (the raw demo's files hold the same positions). With By name (10-5 system) the same names are placed from the template (32 of 32 channels placed: 32 by name (10-5 system).) and move a little towards the centre of the drawing: the demo file puts the outer ring (Fp1 … F7, T7, P7 … O1) on the head line, the 10-5 system one ring (10%) above it.
• Simulated: P1 +2 µV at 60 ms (Oz), N1 −5 µV at 100 ms (Cz), P300 at 350 ms (Pz): Target 10 > Novel 6 > Standard 2 µV, 10 Hz alpha over O1 / Oz / O2 in random phase (it averages out of the ERPs), and after Target the alpha amplitude halves from 350 to 650 ms (power −75%, −6 dB: the alpha decrease, or desynchronization, of step 7), noise. Participants differ by about ±10% in amplitude and ±10 ms in latency.
• ERPs at Pz (grand average): the three lines split after about 250 ms, Target highest, peaking near 350 ms (about 7 µV).
• Mean amplitude, 300–400 ms, Pz: about 6 µV Target, 3.5 µV Novel, 1 µV Standard in every participant. These are lower than the simulated peaks because the window also takes the flanks of the wave, and the average reference takes a little away from every channel.
• Scalp maps (grand average, 300 to 400 ms): the Target map is red around Pz (about +6 µV there), Novel weaker (about +3.5 µV) and Standard almost flat (about +1 µV); Target minus Standard is largest at Pz (about +5 µV). Frontal sites are slightly blue (about −2 µV for Target): with an average reference, a positive focus comes with negative values elsewhere. At 100 ms (type 100 to 100) every map is blue around Cz (the N1, about −3 µV).
• Compare conditions, parametric: repeated-measures ANOVA, p < 0.0001, every pair differs (Target − Standard about +5 µV). Nonparametric: Friedman χ²(2) = 16, p = 0.0003.
• N1: peak amplitude, negative, 50–150 ms at Cz: about −4.5 µV near 100 ms in every condition.
• Checks (after Show ERPs and Measure): no warnings. Trials per condition is a Check: Target and Novel keep about 12 to 15 trials each (15 before the 5 trials rejected in EEGLAB), fewer than 20: enough for the large P300, not for small components. Rejection balance is a Note (the trials were rejected in EEGLAB, so the share of each condition is not known here); Condition balance is OK with the mean amplitude (Standard has about 3 times as many trials as Target or Novel, as an oddball design intends) and a Check with the N1 peak amplitude; Bad channels OK (none); Interpolated channels OK: T7 was interpolated (in EEGLAB) in 8 participants; not among the measured channels (Pz).
• Compare conditions then adds the checks of the test below those rows (all OK on the demo): Sample size (n = 8 participants with a value in every condition), Normality (Shapiro–Wilk on the residuals, p = 0.134), Sphericity (Mauchly's W = 0.998, p = 0.994; the Greenhouse–Geisser p is reported when it is rejected) and the Robustness check (the ANOVA, p < 0.0001, and the Friedman test, p = 0.0003, agree). A new measure removes them with the old test.
• Faults demo (demo_eeg_faults.mat in the demo folder, or eeg/faults/sub-01_faults.set, written with the other EEG demo files: load it with Load EEG files…; one participant, Standard 40, Target 15 and Novel 15 trials): Suggest (step 2) gives 8 bad channels (FC6, T7, T8, TP9 and TP10 noisy, F8, PO9 and PO10 flat); tick Reject trials with 100 µV and click Apply rejection: the 12 trials with a blink go, 9 of them Target. The Checks tab gives 4 warnings: Trials per condition (Target 6), Rejection balance (60% of Target trials but 5% of Standard), Bad channels (8 of 32, 25%) and, measuring at Pz, Interpolated channels (Pz, the measured channel, was interpolated (in EEGLAB): the file's history says so). Condition balance is OK with the mean amplitude (Standard 38 vs Target 6). Without the bad channels marked, every trial exceeds 100 µV (a message says so).
• Time–frequency (step 7; 4 to 40 Hz, 3 cycles, baseline −200 to 0 ms, Channels Oz, grand average): Target turns blue at 10 Hz from about 350 to 650 ms (about −4.5 dB from 400 to 600 ms: the simulated −6 dB, diluted by the noise); Standard and Novel stay near 0 dB (within ±0.2 dB); Target minus Standard is about −4.6 dB there. Below 8 Hz the images are grey: in trials that start 200 ms before the event, no 3-cycle wavelet below 8 Hz fits inside the baseline (the line under the button says so); the trial edges are grey too.
• Phase locking (ITPC) at Cz: about 0.65 to 0.75 at 10 Hz near 100 ms in every condition (the N1 has the same phase in every trial), near chance from 400 to 600 ms (about 0.16 for Standard, 0.25 for Target and Novel; the line above the images gives the chance level: 0.15 and 0.24, higher with fewer trials).
• Band power (Alpha, Oz): Target about −60% from 400 to 600 ms; Standard and Novel near 0 (within about 10%). With these 1 s trials the alpha band has only −20 to 0 ms of baseline (where the 8 Hz wavelet fits), so it is noisy; the line under the button says so.
• Raw demo (Try raw demo (continuous, not cleaned)): the same oddball design as 3 raw recordings (BrainVision Recorder files, 500 Hz, about 100 s each), recorded against FCz (not in the data), markers S 1 / S 2 / S 3 = Standard (40) / Target (15) / Novel (15). The demo fills in the settings below; nothing runs until you click.
• Raw demo, step 2: Suggest gives T7 for every participant (white noise of 150 µV). The files also hold electrode offsets (up to ±400 µV) and slow drift (the 0.1 Hz high-pass removes them), 50 Hz line noise of 5–15 µV (the 30 Hz low-pass, or the 50 Hz notch, removes it) and blinks of about 150 µV at Fp1 / Fp2. With Reference: Average, the N1 at Cz grows from about −2 µV against FCz (Cz is close to FCz) to about −4 to −5 µV.
• Raw demo, step 3: S 1 = Standard, S 2 = Target, S 3 = Novel, −200 to 800 ms, Peak-to-peak 100 µV: exactly the 8 trials with a blink are rejected in each participant (62 trials left); the 6 blinks between trials do not matter. Without marking T7 bad, every trial exceeds 100 µV (a message says so). Then steps 4 and 5 as for the cleaned demo: P300 at Pz, Target > Novel > Standard.
• Raw demo, longer trials for step 7 (cleaned as above, cut from −600 to 1000 ms with the same 100 µV rejection: 52 or 53 of 70 trials are kept, because blinks between trials now fall inside the longer trials): with the baseline −500 to −100 ms every frequency from 4 Hz has values and a baseline; at Oz the alpha band power falls by about 60% after Target from 400 to 600 ms (about −4.3 dB at 10 Hz), Standard and Novel stay near 0 (within about 15%).
• Rodent (continuous, 4 skull screws at 1000 Hz, 60 s, 30 light flashes): cut from −100 to 400 ms gives 30 trials; at V1 a negative peak of about −40 µV at 50 ms and a positive one of about +25 µV at 100 ms, about 30% of that over M1.
• Rodent layout: the EEGLAB, FieldTrip and EEG-BIDS files of eeg/rodent hold the 4 skull screws in mm from bregma: 4 of 4 channels placed: 4 from the file, drawn on the skull: M1-L and M1-R at AP +1 mm, ML −1.5 and +1.5 mm; V1-L and V1-R at AP −3.5 mm, ML −2.5 and +2.5 mm. The EDF, BDF, XDF, BrainVision and plain .mat versions have no positions (0 of 4 channels placed: the screw names are not 10-5 names); a .csv with name, ap, ml loaded with Positions file… places them.
• Rodent scalp map at 50 to 50 ms: a flat map between the four screws, about −40 µV over V1 and −12 µV over M1 (30% of it), changing smoothly from front to back; nothing is drawn outside the screws.