Neurofeedback and BCI on consumer headsets
For: Developers building on four- to eight-channel wireless headsets. 15 steps · ~17 h.
Why this order
This path assumes you have a headset, or will, and want to build something on it that a stranger could check. It is ordered so that each lesson removes one way the thing you build could be wrong without your knowing: first about the signal, then about the stream, then about the loop, then about the claim.
What the hardware can deliver (L0.3–L0.5). Sampling rate, hardware bandwidth and the difference between the two set the ceiling on everything downstream, and on a wearable device the output rate and the usable bandwidth are routinely confused. L0.4 and L0.5 build the ability to look at a live trace and say what is wrong with it — the only debugging tool an online system has, and the one the quality gate of L7.12 is a port of.
Filtering, where causality becomes concrete (L1.5, L1.6). Online there is no future, so the filter delays the signal and the delay is a row in the budget. L1.6 matters because a mains line is the normal state of an un-notched consumer recording, and what a device does about it before the file exists decides what you can still see.
Cleaning that can run causally (L2.7). What is left when ICA is not an option, and why.
Then the headsets themselves (L7.5, L7.10, L7.11). L7.5 is the neutral survey — channel count decides which analyses exist; rate is not bandwidth — and L7.10 makes it concrete for two four-electrode geometries, frontotemporal-4 and occipitotemporal-4, from documented facts rather than from any recording. L7.11 is the stream: one interface for replay, synthetic and device; timestamps, dropped samples and drift measured rather than assumed; a session recorded in the site’s own format and read back.
The loop (L7.3, L7.12, L7.13). L7.3 is the general lesson on real-time processing and the latency budget; it sits here, after the stream, because its budget has rows the stream lessons define. L7.12 puts the gate first, on the raw block, and shows why. L7.13 states a feedback signal in full and measures what smoothing costs.
The claim (L7.14, L7.15, L7.16). A protocol that can fail: sham modes, blinding, pre-registration and the valid learning test against the naive one that flags eight of twelve replayed minutes. A first decoder on four channels, calibrated once, frozen and applied, with every accuracy beside its chance band. Then the application, packaged, probed, gated, sealed, reported — and a hardware run documented with no efficacy claim.
The path skips the ERP and inference chains almost entirely; the entries below say which of them are worth returning for. Nothing in it names a device except in a dated fact table, once per lesson; the numbers come from ds-eegbci channel subsets and the library’s synthetic stream, never from a headset.
Paths do not add content; they filter and order the ladder (§7). Lesson pages keep their own prerequisite links.
Lessons (15; about 16 h 40 min of lesson time)
- L0.3 Amplifiers, sampling and recording · 45 min
- L0.4 Reading raw traces · 60 min
- L0.5 The artifact atlas · 75 min
- L1.5 Filters: FIR, IIR, and what they do to your data · 75 min
- L1.6 Line noise · 40 min
- L2.7 ASR, SSP and alternatives · 45 min
- L7.5 Mobile and consumer EEG · 60 min
- L7.10 What a four-channel headband measures · 60 min
- L7.11 Getting a stream: timestamps, dropped samples, drift, recording · 75 min
- L7.3 Real-time processing and neurofeedback · 90 min
- L7.12 Signal quality you can act on live · 60 min
- L7.13 The feedback signal: band power in a buffer · 75 min
- L7.14 A protocol that can fail: sham, blinding, pre-registration, the valid test · 60 min
- L7.15 A first BCI on four channels · 90 min
- L7.16 Build the application · 90 min
Skipped prerequisites (optional reading)
Each of these is a prerequisite of a lesson this path includes. They are optional here, not removed — the lesson pages keep their prerequisite links.
- L0.2 Electrodes, montages and the 10-20 system — a prerequisite of L0.3. Read it before L7.10 if the names TP9, AF7, O1 and T7 do not yet place themselves on a head.
- L1.2 Time and frequency domains — a prerequisite of L1.3; the Fourier grounding under every band-power feature.
- L1.3 Power spectral density — a prerequisite of L1.5, L7.10 and L7.13, and where the variance of a short-window power estimate is explained — the reason an online band-power readout is noisy.
- L2.5 Artifact rejection strategies — a prerequisite of L7.15; the epoch-rejection recipe the P300 section reuses.
- L2.6 EOG regression and ICA — a prerequisite of L2.7. Worth reading for the comparison: it is what the causal alternatives are alternatives to, and L7.12 measures what ICA at rank four does and does not do.
- L4.3 Multitaper and filter-Hilbert — a prerequisite of L7.3; band amplitude with a known latency, which is one of the three derivations L7.13 compares.
- L4.4 Baseline normalization and ERD/ERS — a prerequisite of L7.13; where a baseline is defined and what a feedback value is relative to.
- L6.5 Decoding as inference — a prerequisite of L7.15; the chance band and the leakage catalogue that lesson holds its accuracies to.
- L6.6 Reporting standards — a prerequisite of L7.14; the checklist the CRED-nf items belong to, easier to accept before you have a system to defend.
- L6.4 Circularity and analytic flexibility — a prerequisite of L7.14, and the standard failure mode of neurofeedback evaluation that pre-registration exists to close.
Lessons in order
- L0.3 · Amplifiers, sampling and recording ~45 min
- L0.4 · Reading raw traces ~60 min
- L0.5 · The artifact atlas ~75 min
- L1.5 · Filters: FIR, IIR, and what they do to your data ~75 min
- L1.6 · Line noise ~40 min
- L2.7 · ASR, SSP and alternatives ~45 min
- L7.5 · Mobile and consumer EEG ~60 min
- L7.10 · What a four-channel headband measures ~60 min
- L7.11 · Getting a stream: timestamps, dropped samples, drift, recording ~75 min
- L7.3 · Real-time processing and neurofeedback ~90 min
- L7.12 · Signal quality you can act on live ~60 min
- L7.13 · The feedback signal: band power in a buffer ~75 min
- L7.14 · A protocol that can fail: sham, blinding, pre-registration, the valid test ~60 min
- L7.15 · A first BCI on four channels ~90 min
- L7.16 · Build the application ~90 min