Mobile and consumer EEG

Objectives

• Characterize dry and low-channel-count systems
• Identify motion artifacts
• Determine which analyses survive with 4–8 channels
• Validate against a research system

Why this matters

A headset that streams 128 samples per second and reports “focus” is not a small research amplifier; it is a different instrument, and the difference is not a matter of quality but of what the numbers in the file are. This lesson characterises three device classes from the facts their own datasets record — channel count, electrode type, reference, hardware bandwidth, what the vendor computed for you — and asks the only question that matters for an analysis plan: given this file, which results are still available?

Concepts

Three classes, described from what their datasets record

Everything in this table comes from the dataset entries in this site’s directory. It ranks nothing and recommends nothing, and it names devices once, as a fact, because the acquisition determines what the data can support.

	ds-eegbci	ds-brain-invaders	ds-areeg
Class	research cap	research dry	consumer
Electrodes	64, gel, 10-10 cap	16 active dry (g.Sahara)	14 wireless saline
Sampling rate	160 Hz	512 Hz	128 Hz nominal
Online filters	none	none	device factory front-end; not documented in the paper
Reference	TODO(confirm) in the catalog	right earlobe	CMS/DRL above the ears
Mains	60 Hz	50 Hz	TODO(confirm)
Participants	109	64 released of 71 recorded	22
Licence	ODC-By 1.0, open	CC BY 4.0, open	CC BY 4.0, open

Read that table as three sets of constraints rather than three products. Sixty-four gel electrodes buy spatial sampling and a decomposition that has somewhere to go; sixteen dry electrodes remove the gel and the preparation time and raise the contact impedance; fourteen saline electrodes on a consumer frame give a wearable recording with a fixed, undocumented front end. Each buys something and each gives something up, and the honest way to describe one is to say which.

The sampling rate is not the bandwidth

The Nyquist frequency of a 128 Hz stream is 64 Hz. The usable bandwidth of the consumer headset behind ds-areeg, ds-mpeng, ds-vep and ds-emotions ends near 43 Hz: the directory records, for all four, a nominal 128 or 256 Hz output with an ≈ 43 Hz hardware bandwidth and built-in 50/60 Hz notches. The file will happily give you a number for 50 Hz power. It is not a measurement of the head.

This is pf-hardware-bandwidth-ceiling, and it has a second edge that is easy to miss. A built-in notch removes the diagnostic as well as the noise. On ds-eegbci, which has no hardware filters at all, line noise is visible, so you can see whether an electrode was badly connected, whether the environment changed between subjects, and whether a notch you applied did what you intended (L1.6). On a device that notches before you see the data, the one channel-quality signal that is always present has already been deleted, and you cannot recover it.

Two consequences for an analysis plan. Anything above roughly 40 Hz is off the table on these devices — not “noisy”, not “needs more trials”, unavailable. And the aperiodic fit range (L1.7) must end below the ceiling and exclude the notch neighbourhood, or the exponent you report is a description of the device’s filter.

Vendor-derived columns are not measurements you made

ds-mpeng’s exported CSV interleaves its 14 EEG columns with contact-quality, “performance-metric” and band-power columns that are vendor-derived, not raw signal. They look like data. They are the output of an undocumented pipeline running inside a device, and three things follow:

• You cannot state their provenance in a methods section. You do not know the filter, the window, the band edges, the normalisation or the update rate, so you cannot say what was computed, and a reader cannot reproduce it.
• They are not independent of each other or of the EEG columns. A band-power column and a “performance metric” derived from it are the same measurement twice.
• They may already encode a rejection. If the device suppresses output when contact is poor, the column is conditioned on a quality criterion you did not choose.

The same dataset shows what that costs. Only about 50 % of its samples meet the authors’ own quality criterion — a vendor-computed EQ.OVERALL at or above 75 % — during intense play. That is the authors reporting, honestly, that half of the recording under the condition of interest fails the quality gate their own device supplies. It is a fact about a real recording in the field, and it is the number to hold in mind when a study proposes to record cognition during movement.

What motion does to the signal, and why these systems see more of it

Movement enters the recording through mechanisms that have nothing to do with the brain and everything to do with mechanics: an electrode shifting against skin changes the half-cell potential at the interface, which appears as a step or a slow swing far larger than any cortical signal; a cable swinging changes capacitance; and a subject who is moving is also a subject whose neck, jaw and face muscles are active, which is pf-muscle-as-gamma.

The reason a low-channel dry or saline system shows more of this is a chain of ordinary physics, and the pitfall entry states it: higher contact impedance and looser mechanical coupling pick up more muscle and more movement. That is a statement about electrode-skin contact, not about a manufacturer. It also means the artifacts and the signal are correlated with the condition in exactly the studies these devices are bought for — ambulatory, gaming, sport, the field — so an effect of “engagement” and an effect of moving more while engaged are not separable by any statistic applied after the fact.

Two habits follow. Record a movement reference — accelerometer, video, task log, a channel that is deliberately not brain — so that motion is a measured covariate rather than an inference. And decide before the study which frequency ranges you will interpret, given that EMG overlaps everything above about 20 Hz and the device’s own ceiling may sit at 43.

Channel count decides which analyses exist

Channel count is not signal quality. It is the rank of your data and the resolution of your map, and it removes whole classes of analysis rather than degrading them smoothly:

• Single-channel spectral measures survive: a posterior alpha peak, the individual alpha frequency, the aperiodic exponent and offset over a stated range. One well-placed electrode is enough, and adding channels mostly buys robustness to one of them failing.
• Contrasts at a named site survive if the site is present. A P300 needs a midline parietal electrode; mu ERD lateralisation needs C3 and C4 (L4.4). A montage without them does not measure the effect weakly, it does not measure it.
• Anything that decomposes the channel space fails first. ICA can return at most as many components as the data have rank (L2.6), and rank falls further with re-referencing and interpolation (pf-interpolation-rank). At four channels there is no decomposition that separates blink from frontal theta; there are four mixtures and four sources.
• Anything spatial fails: interpolated topographies, the surface Laplacian, source analysis, and any connectivity measure whose validity rests on a spatial model (L5.1).

The rule of thumb to carry away is that low channel counts remove the spatial analyses and the rank-based ones, and leave the single-channel spectral ones, and that the bandwidth ceiling is a separate restriction which removes frequency ranges on any montage.

Validating against a research system

“It shows alpha” is not a validation. Eyes-closed posterior alpha is the easiest signal in the field and a wire near a head will find something that looks like it. A validation design has to be able to fail:

1. Record both systems on the same heads, on the same task, simultaneously where the hardware allows it and back to back in counterbalanced order where it does not. Anything else confounds device with session, subject and time — pf-site-device-confound in its simplest form.
2. Fix the measure and the pipeline in advance, and run the same analysis on both, restricted to the passband both systems share. Comparing a 43 Hz-limited device against a full-bandwidth one over 1–100 Hz measures the filter.
3. Report agreement per subject with an interval, not a group correlation. A correlation across subjects can be high while the per-subject values disagree by more than the effect you intend to study; the quantity that matters is the difference between the two devices in one person, and its spread.
4. State what the comparison cannot settle. Two systems can agree on IAF and disagree on everything above 30 Hz, and a validation of one measure is not a validation of the device.
5. Report the failures: the electrodes that never made contact, the sessions discarded, the participants whose hair defeated the dry pins. A validation that reports only the recordings that worked has measured its own inclusion criterion.

Explore

This lesson has no widget of its own; two from earlier levels do the work, and the point is to arrive at them with a device in mind.

• Open the sampling explorer in its math mode and set the output rate to 128 Hz. Note where Nyquist lands (64 Hz) and then mark 43 Hz on the same axis. The gap between the two is the band the file reports and the hardware never delivered.
• Open the filter sandbox in its notch mode on the line-noise trace. Watch what a notch does to a spectrum, then ask what you would have seen if that notch had been applied inside the amplifier before the file existed — and what you would no longer be able to check.
• On paper, take the last analysis you ran and list, for each step, the minimum number of electrodes and the minimum bandwidth it needs. Most plans have never been written down this way, and most have one step that is doing all the constraining.

Practice

What survives at fourteen channels: three device classes measured with one estimator, the hardware bandwidth ceiling behind a nominal sampling rate, the vendor-derived columns that are not signal, and a controlled degradation of one recording from 64 electrodes to four nb-7-5-low-channel

Level 7 ~5 min

notebooks/L7/nb-7-5-low-channel.ipynb

Downloads from ds-eegbci, ds-brain-invaders, ds-areeg.

Open in Colab Download Read it here

The notebook runs the same measures across device classes: a 64-channel research cap (ds-eegbci), a 16-channel dry research headset (ds-brain-invaders), and a 14-channel consumer headset (ds-areeg) — each analysed where its paradigm supports the measure — and then the controlled degradation, a 4-channel subset of ds-eegbci, where only the channel count changes. ds-mpeng is named and not loaded: its licence is non-commercial and consent-limited, so no notebook downloads it and no asset derives from it.

Exercises

Exercise ex-7-5-feasible-at-four

Multiple select

You have a 4-channel recording at Fp1, Fp2, O1, O2 from a research amplifier with a flat response across the whole EEG range, so that this question is about channel count alone. Which of these five analyses remain feasible?

Options (select all that apply)

Estimate the individual alpha frequency from an eyes-closed segment.Fit the aperiodic exponent and offset of the spectrum on a posterior channel.Measure mu-rhythm ERD lateralization over sensorimotor cortex during a hand movement.Run ICA and remove a blink component while keeping frontal theta.Interpolate a scalp topography of the P300 and report where it is maximal.

Check

Hint

Exercise ex-7-5-nominal-vs-usable

Numeric

A consumer headset streams 128 samples per second. The directory records its hardware bandwidth as ending near 43 Hz. What is the Nyquist frequency of that stream?

Your answer Hz

Exact answer required, in Hz.

Check

Hint

Exercise ex-7-5-vendor-column

Multiple choice

A consumer export gives you 14 EEG columns, a per-electrode contact-quality column, a set of band-power columns and a vendor 'performance metric'. Your methods section has to state what each reported quantity is. Which statement is defensible?

Options

Report the vendor band-power columns as band power: they are labelled with the band edges, so the computation is documented.Report analyses computed from the 14 EEG columns, state the vendor columns as device outputs of an undocumented pipeline where they are used at all, and do not treat them as independent measurements.Report the performance metric as the primary outcome, since it is the quantity the device was validated to produce.Drop the EEG columns and use only the vendor columns, which are already cleaned and therefore need no preprocessing.

Check

Hint

Exercise ex-7-5-iaf-4ch

Numeric

From the notebook's controlled degradation: taking the 64-channel estimate as the reference, what is the median absolute difference in individual alpha frequency when the same recordings are re-analysed on the 4-channel subset?

Your answer Hz

No answer key yet — work it out and compare with the notebook.

Check

Hint

Exercise ex-7-5-validation-design

Free response

A team proposes to validate a 4-channel consumer headset against a 64-channel research cap by recording 20 people at rest on each device on different days, correlating alpha power across the 20 participants, and concluding from r = 0.82 that the headset is suitable for field studies. Rewrite the design, and state what the rewritten study could and could not conclude.

Your answer

Reveal model answer

Pitfalls

Pitfall

EMG reported as gamma

Symptom

Broadband high-frequency power over temporal/frontal sites tied to jaw or neck.

Cause

Scalp muscles (temporalis, frontalis, masseter, the neck extensors) produce electrical activity that is large compared with cortical gamma, broadband from roughly 20 Hz upward with no single peak, and sits directly under the electrodes that “show the effect”. Its amplitude is modulated by anything that changes muscle tone — and many experimental manipulations do. The spectrum of EMG overlaps the…

Detect

• Topography: cortical gamma from a focal source is rarely largest at the edge of the montage; EMG is. Look at T7/T8, F7/F8, FT sites and the occipital rim. - Spectral shape: EMG raises the floor across a wide range with no peak; a genuine oscillation has a peak above the aperiodic background (L1.7, L4.6). - Time course: EMG turns on and off with muscle events (swallows, jaw movement, blinks with…

Fix

• Instruct and monitor: relaxed jaw, no talking, breaks; record a facial EMG channel where high frequencies matter. - Clean before measuring: ICA with muscle-component removal, or rejection of segments with high-frequency power above a threshold (L2.5, L2.6); state what was removed. - Report a peak, not a band: use spectral parameterization to show a gamma peak above the aperiodic background befo…

Full entry with example →

Pitfall

Usable bandwidth far below the Nyquist frequency

Symptom

"Gamma" analysed at 128 Hz output from a headset whose hardware rolls off near 43 Hz; a 500 Hz recording with a ~30 Hz hardware low-pass.

Cause

The sampling rate sets the Nyquist frequency (half the rate), which is the highest frequency the file could represent. It says nothing about the highest frequency the amplifier passed. Anti-alias and other hardware low-pass filters sit well below Nyquist by design; consumer devices decimate from a high internal rate after a low-pass that ends far below the output rate. The band between the hardwa…

Detect

• Read the device specification and the dataset descriptor for hardware bandwidth, not just the sampling rate. - Compute a PSD up to Nyquist on a quiet segment: a knee after which the spectrum falls steeply, then flattens onto a floor, marks the ceiling. The knee is at the same frequency in every channel and every subject. - Check that the spectrum above the knee shows no physiology: no reactivit…

Fix

• Restrict every analysis to the hardware passband and state that passband in the methods. - For decimation and resampling, the anti-alias filter must sit below the new Nyquist frequency (L1.1) — and the usable band is still bounded by the hardware ceiling, whichever is lower. - When a research question needs a band the device cannot deliver, change the device, not the analysis. - Teach it as a f…

Full entry with example →

Reading

1. Gramann et al. (2014). Mobile brain/body imaging. unverified