What a four-channel headband measures
A four-electrode headband described from documented facts — which four sites, what rate, what bandwidth, what reference, what the device filtered before you saw the stream — and the analyses that follow from them, for the two geometries the consumer path names.
Prerequisites: L0.3 · Amplifiers, sampling and recording, L0.5 · The artifact atlas, L1.3 · Power spectral density, L7.5 · Mobile and consumer EEG
10 claims on this page are unverified. TODO(confirm) marks a specific statement the author has not yet checked against a
primary source. Everything else on this page has been reviewed. Treat a marked claim as
provisional and go to the cited source rather than quoting the sentence.
What you will be able to do
- State the montage, rate, bandwidth and reference of a headset from documented facts
- Predict which analyses a frontotemporal versus an occipitotemporal four-electrode set supports
- Identify what an on-device notch or 1–40 Hz filter removes from what you can later check
- Read a first-look report of a four-channel stream
Why this matters
Everything the rest of this path builds — a stream, a quality gate, a feedback loop, a decoder, an application — runs on four electrodes. Four is not a small sixty-four. It is a different instrument, and the difference is a short list of facts from which the analyses that remain available follow: which four sites, at what rate, with what bandwidth, against what reference, and what the device did to the signal before it reached you. L7.5 made this argument for three device classes from their datasets. This lesson makes it for the two four-electrode geometries the consumer path names, from the facts their drivers and documentation record, and then measures what those facts imply on real data where an honest replica exists and on synthetic data where it does not.
The habit to take from it is the one L7.5’s channel-count section taught: describe, never rank. A frontotemporal headband and an occipitotemporal one answer different questions, and the table that says which is a table of sites and bands, not a verdict.
Concepts
Two geometries, named by what they are
The path describes a device by its montage class — the geometry of its electrodes, named the way positions.json names every other montage on this site, by source rather than brand:
frontotemporal-4: TP9, AF7, AF8, TP10. Two sites behind the ears over posterior temporal cortex, two on the forehead over the frontal poles. The forehead pair sees blinks and frontal muscle first; the pair behind the ears sees a weakened posterior alpha and the temporalis.occipitotemporal-4: O1, O2, T7, T8. Two sites over the occipital poles, two over the temporal lobes at the level of the ear. The occipital pair sees posterior alpha and a flicker response at full strength; nothing in the set sees a blink as a blink.
Neither includes Cz, Pz, C3 or C4. That single fact settles two of the questions people most often bring to a headband, and the matrix below says how.
What this lesson connects to
The device behind each geometry is named once here, as a set of verified facts with a date, in the shape L7.5 uses. Everywhere else this path says “the driver’s board table” and describes the device by its class. Access facts are stated the way dataset access classes are — as constraints you need to know, not as reasons to prefer a device.
| frontotemporal-4 | occipitotemporal-4 | |
|---|---|---|
| Device (named once) | Muse 2 and Muse S (InteraXon); the Muse S Athena firmware is a third driver id | BrainBit (BrainBit) |
| Electrodes | TP9, AF7, AF8, TP10, plus an AUX input | O1, O2, T3, T4 — the driver uses the older names; T3/T4 are T7/T8 (L0.2) |
| Sampling rate | 256 Hz | 250 Hz |
| Documented band | TODO(confirm) | 0–100 Hz |
| Reference | TODO(confirm) as the driver reports it | TODO(confirm) |
| On-device filters | TODO(confirm): none known on the raw-EEG preset; the companion app’s stream is not this stream | an on-device notch and a 1–40 Hz band-pass are optional and must be off for this course; the driver switch and its default are TODO(confirm) |
| Transport | native Bluetooth LE, no dongle | native Bluetooth LE, no dongle |
| Driver route | the driver’s board table (three ids; driver 5.22.0) | the driver’s board table |
| Raw EEG access | free, through the open-source driver | free, through the open-source driver |
| Contact-quality exposure | TODO(confirm) | TODO(confirm) |
| Verified | 2026-09-20 | 2026-09-20 |
Why these two come first in the path is a fact stated once: each reaches Python through a free open-source driver over native Bluetooth with no dongle (TODO(confirm): cite the driver’s licence page for “open-source”). Two further headsets the author owns are named, with their access facts and a verification date, in the library’s hardware notes (loop/HARDWARE.md, “Not settled”) — raw EEG on one requires a paid developer licence, on another it is free — and neither enters this track until its route is verified.
Every row marked TODO(confirm) is a fact the library’s acceptance report (eegloop check, L7.11) is built to settle from a live stream: the measured rate against the claimed one, the bandwidth edge, the timestamp structure, the units. The report a human files after running it on the device is what turns the marker into a number, and no number from it is quoted on a page.
An emulation that is exact in geometry, and one that does not exist
ds-eegbci is recorded on a 64-channel gel cap at 160 Hz with no hardware filters. Its cap has O1, O2, T7 and T8, so the occipitotemporal geometry can be cut out of it exactly — the same four sites, on the same head, in the same minute. The notebook does this for twelve subjects, and every number below that says “on the replica” is a number from those four channels of a research amplifier. It is a replica of the geometry and of nothing else: the electrodes, the contact, the reference and the front-end are a gel research system’s, and a dry headband at the same sites is a different measurement whose facts belong in its own sidecar.
The cap has AF7 and AF8 but no TP9 or TP10. The notebook computes, rather than asserts, which of its sites are nearest to the two it lacks — T9 and T10, then TP7 and TP8, by distance on the unit sphere — and each of those is a different location, not the same electrode under another name. So the frontotemporal geometry has no honest public emulation on this site. It runs on the library’s synthetic stream, which carries the real channel names of that geometry with planted answers, and every figure and number from it says so. That stays true until a licensed recording of the geometry lands (the directory holds one candidate, ds-cognitive-states-students, read from its record as CC BY 4.0; nothing derives from it until the author adds its registry entry, §10.1 rule 3) or until the author’s own recordings meet the conditions §10.6 sets.
The bandwidth is the band, not the Nyquist frequency
A 250 Hz stream has a Nyquist frequency of 125 Hz. A documented band of 0–100 Hz says that the highest frequency you may analyse is 100 Hz, and the ceiling is the band. L0.3, L7.5’s sampling-rate section and pf-hardware-bandwidth-ceiling made this point about the rate; the notebook makes it as a measurement, with the library’s bandwidth edge — the frequency above which the spectrum stays 20 dB below its 4–20 Hz level for the rest of the way to Nyquist. On the replica, on a cap documented as having no hardware filters, an edge is detectable on only about half the twenty-four occipital channel-runs across the twelve subjects, and where it exists it sits anywhere from about 40 to 75 Hz rather than tightly at 40–60; on the temporal pair it is absent in all but two of the twenty-four: the amplifier’s own roll-off and the signal’s own 1/f fall are already there before any device filter is. On a stream that arrives band-limited by the device, the edge sits at the filter, and the report says which it found.
What an on-device notch removes from what you can later check
The optional filters one of the two devices offers — a notch at the mains and a 1–40 Hz band-pass — remove the mains line and the high band. That is not their cost. Their cost is the diagnostics. The live quality gate (L7.12) reads three things from the raw block that a band-pass deletes: the power at the mains against its neighbourhood, which tells a coupled electrode from a clean one; the 25–100 Hz band, which tells muscle from EEG; and the DC level and spread, which tell a railed or disconnected channel. The notebook applies the same two operations offline, at the replica’s own mains frequency, and runs the gate before and after: the line ratio collapses to one on every channel, the bandwidth edge moves from the amplifier to the filter, and the fraction of windows the gate labels drops toward zero — not because the channels became clean but because the gate can no longer see. L7.5’s sentence about the notch that deletes the diagnostic as well as the noise is this, measured.
The rule that follows is the one the library’s board table records for that device: on-device filters off for this course, and if a stream arrives filtered, the sidecar says so and the analysis says which checks it could not run.
Reading a first-look report of a four-channel stream
L0.6’s first look does not change at four channels; what changes is what the rows can promise. The sampling-rate row compares the header with the documented rate. The channel-count row says four of four, and says nothing about whether four is enough. The header’s low-pass value is what the file says, not what the amplifier passed. Add the four things a live stream can report and a file cannot — the measured rate against the claimed one, the samples lost, the structure of the timestamps, the package counter’s accounting — and you have the acceptance report eegloop check prints for any source and L7.11 reads in full. On the synthetic frontotemporal stream the notebook prints one; its first line says the stream is synthetic.
Which analyses survive which geometry
The matrix, with the reason in each cell. It ranks nothing.
| Analysis | frontotemporal-4 | occipitotemporal-4 |
|---|---|---|
| Eyes-closed alpha peak, individual alpha frequency | yes — posterior temporal sites see it, weakened (synthetic only here; TODO(confirm) on a licensed recording) | yes — O1/O2: every one of twelve subjects on the replica shows a peak 3 dB above trend |
| Alpha reactivity, eyes closed against open | yes — the test of whether a channel records EEG at all | yes — measured on the replica |
| Blink detection | yes — AF7/AF8 are frontal; the gate’s blink rule needs a frontal site | no — the gate says “no frontal channel in this montage” and makes no blink claim |
| Muscle (EMG) detection | yes — 25–100 Hz RMS on any channel; the temporalis sits under TP9/TP10 | yes — the temporalis sits under T7/T8 |
| Line-noise diagnostic | yes, if the stream arrives unfiltered | yes, only if the optional on-device notch is off |
| Steady-state flicker response (SSVEP) | no — no occipital site | yes — O1/O2 (L7.15) |
| P300 | no — no Pz or Cz | no — no Pz or Cz; L7.5’s point about which site the P300 needs |
| Mu desynchronisation, motor imagery | no — no C3/C4 | no — no C3/C4; motor imagery is unavailable on both, and L7.15 says so |
Two consequences shape the rest of the path. A feedback signal on the occipitotemporal geometry can be posterior alpha; on the frontotemporal geometry it is a weaker posterior alpha or a frontal band, and the choice is stated as part of the signal’s definition (L7.13). And the first BCI on four channels (L7.15) is eyes-closed against eyes-open, a flicker response, or a blink switch — never motor imagery, whatever the marketing of any device says.
The data behind this lesson
ds-eegbci (ODC-By 1.0, open): S001–S012, runs R01 (eyes open) and R02 (eyes closed), restricted to O1, O2, T7, T8 — the occipitotemporal-4 geometry cut from the 64-channel cap. The frontotemporal-4 geometry appears only on the library’s synthetic stream, with its real channel names, and is labelled synthetic wherever it is shown. No consumer recording is used: the directory’s only downloadable consumer front-end (ds-areeg) has a fourteen-channel saline montage that is neither geometry, and it enters this path in L7.12 for what its vendor columns show. The pitfall figure pf-contact-is-not-signal places a real eyes-closed trace beside two synthetic ones of the same amplitude.
Explore
Select frontotemporal-4 and occipitotemporal-4 in turn. Then switch to eegbci-64 and find O1, O2, T7 and T8 on the cap — and look for TP9 and TP10, which are not there.
Practice
What a four-channel headband measures: the two geometries as facts, a first-look report of a four-channel stream, the occipitotemporal replica on twelve ds-eegbci subjects, the frontotemporal geometry on synthetic data with its real names, the bandwidth edge measured, what an on-device notch removes from the diagnostics, and which analyses survive which geometry nb-7-10-headband
Downloads from ds-eegbci.
Exercises
Exercise ex-7-10-band-not-nyquist
NumericA headset streams at 250 Hz and its documentation states a band of 0–100 Hz. What is the highest frequency, in hertz, at which you may analyse this stream?
Exercise ex-7-10-replica-alpha
NumericOn the occipitotemporal-4 replica cut from ds-eegbci — O1, O2, T7, T8 of the 64-channel cap — the notebook measures the eyes-closed alpha peak on O1+O2 for subjects S001 to S012 with this site's estimator (largest residual above a straight-line fit of log power against log frequency, counted only when it exceeds 3 dB). How many of the twelve subjects show a peak?
Exercise ex-7-10-what-survives
Multiple selectSelect every statement that is true of the two four-electrode geometries as this lesson describes them.
Pitfalls
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…
Hardware notch or band-limit inside the analysis band
- Symptom
- Spectral hole at 50/60 Hz or a roll-off at 40–50 Hz mistaken for a physiological dip; double-notching in software.
- Cause
Amplifiers and headsets can apply filters before the data are stored: a notch at the mains frequency, a low-pass well below the Nyquist frequency, a high-pass. These are recorded in the file header at best and in the device documentation at worst. Once applied they cannot be undone, and they shape every spectrum computed from the data. An analyst who does not read the hardware fingerprint interpr…
- Detect
- Read the header and the dataset descriptor for online filters before computing anything; record them in the first-look report (L0.6). - Compute a PSD on a long, quiet segment up to the Nyquist frequency and look for: a narrow hole at 50/60 Hz, a knee where the spectrum starts falling faster than the aperiodic trend, and a floor. - Check whether the hole or knee is identical across all channels…
- Fix
- Restrict the analysis band to the hardware passband: fit aperiodic models and report band power only below the low-pass knee and away from the notch. - Do not notch again in software; exclude the notch bins from fits and from any narrow-band measure. - State the hardware filters in the methods, alongside the software filters. - When comparing datasets or devices, compare only within the common…
A green contact indicator is not a signal
- Symptom
- Every electrode's contact indicator is green, and the trace under it is mains, muscle or a railed amplifier.
- Cause
A contact indicator is an impedance measurement: the device drives a small current through the electrode and reports whether the return path is low enough. That says the electrode touches skin. It does not say what the electrode is recording. A dry electrode on hair with a good contact reading picks up whatever is largest at that site — mains coupling through a poor reference, the temporalis and…
- Detect
- Line ratio. Welch power at the mains frequency against the median power 2–8 Hz away. A ratio far above ten on one channel, not shared by its neighbours, is coupling on that electrode, whatever the indicator says. - High-frequency ratio. RMS in 25–100 Hz against RMS in 1–25 Hz. Muscle raises it; EEG does not. Ask the participant to clench and watch which channel answers. - Flatness. A window’s s…
- Fix
- Measure quality from the signal, not from the indicator, and gate feedback on what you measured: a channel labelled unusable freezes the display rather than feeding it. - Reseat the electrode, part the hair, moisten a dry contact; re-check by the signal criteria, not by the colour. - Keep the device’s own filters off if it lets you, so the diagnostics stay checkable; if it does not, record that…
Terms used here
Every term this lesson leans on, with the one-line meaning. Follow a term for its full entry.
- Montage
- The set of channel derivations displayed or analysed; also the standard electrode-position layout.
- Nyquist frequency
- Half the sampling rate; the highest frequency that can be represented without aliasing.
- Impedance
- Opposition to current flow at the electrode–skin interface; lower values reduce noise pickup.
Reading
- Gramann et al. (2014). Mobile brain/body imaging. unverified