EEG as the First Laboratory
EEG is where the abstractions in the main roadmap become physical. It is also the module most easily done badly, because a scalp trace looks interpretable long before you can interpret it.
The bounded core is written into the checklists of Blocks 1 to 6 and counts inside those blocks' stated load. It is not extra work running in parallel.
neuro-roadmap.md (the main route) and templates.md (every form you copy from).
01 The honest endpoint –
Research-level EEG literacy. You can explain what EEG physically measures; distinguish electrode, channel, derivation, reference, and montage, and compute the arithmetic between them; describe a recording in standard terminology without diagnosing it; audit how a dataset's labels were produced; state what a measurement cannot establish; and recognise when expert adjudication is required.
This is not clinical competence, certification, independent reporting authority, or diagnostic authority. Those require formal supervised clinical training and no part-time roadmap should claim them.
Two possible endpoints, depending on your circumstances
| Without a supervisor | With a supervisor and an approved case bank | |
|---|---|---|
| Endpoint | Documented introductory EEG research exposure: accurate vocabulary, montage and reference literacy, artefact and label-provenance reasoning | Research-level literacy supported by adjudicated work on unseen continuous recordings |
| Pattern-recognition claim | None. Accuracy scores are private learning checks | Externally supported within the supervisor's stated scope |
If no supervisor and no case bank exists, continue on the atlas route, remove all sign-off language, and make no pattern-recognition claim anywhere in your outputs. That is the base case, not the exception.
02 What is core and what is conditional –
2.1 Core, scheduled inside Blocks 1 to 6
Kept whatever Gate A decides.
2.2 Conditional, only if you stay with EEG after the Interest Gate
If Gate A takes you elsewhere, you keep the core and drop the conditional. Nothing is wasted: the core is measurement literacy, and it transfers to any modality.
03 Sources –
3.1 Primary atlas, free
Britton JW, Frey LC, Hopp JL, et al.; St. Louis EK, Frey LC, editors. Electroencephalography (EEG): An Introductory Text and Atlas of Normal and Abnormal Findings in Adults, Children, and Infants. Chicago: American Epilepsy Society; 2016. On NCBI Bookshelf.
| Chapter | Use it for |
|---|---|
| Orderly approach to interpretation | The description sequence, which is the backbone of this module |
| The normal EEG | Posterior rhythm, reactivity, sleep transients, age dependence |
| Benign variants | Section 6 below |
| Common artifacts | What is not brain |
| EEG in the epilepsies | Discharge morphology and its limits |
3.2 Standards, tools, and free signal processing
Standards. ACNS guidelines on electrode nomenclature and montages. ILAE 2025 seizure classification (Beniczky S, Trinka E, Wirrell E, et al. Epilepsia. 2025), with the practical guide.
On ILAE versions. Learn 2025 as current and know 2017 as well, because most literature you will audit was written under it. Fluency in both is the correct position for anyone reading a decade of papers.
On ACNS critical-care terminology. Learn it only if your capstone is a critical-care question. It is not the native vocabulary of teaching corpora. Learn your corpus's own labels and write the harmonisation map instead.
Tools. MNE-Python tutorials, particularly sensor locations.
Signal processing, free. Mike X Cohen has 100+ hours of free lectures, which are optional lookup material rather than a hidden compulsory curriculum: consult the specific lecture your current task needs. He is a former neuroscience professor and author of Analyzing Neural Time Series Data (MIT Press), the standard reference in this area. Use the free material in Blocks 2 to 5 for intuition about filtering, frequency content, and time-frequency representation.
3.3 The one paid purchase, conditional on Gate A
| If Gate A sends you | Then |
|---|---|
| To an EEG or signal-based capstone and the Block 6 pilot exposed a demonstrated gap in filtering, spectral, or time-frequency reasoning | Buy Cohen's signal processing course as the learning contract's method resource, not stacked on top of it. Roughly 12 hours, has Python code paths |
| Anywhere else, or no demonstrated gap | Do not buy it. The free lectures cover what the core module needs |
Why it earns the money in the EEG case: from Block 6 onward you set filter cutoffs, epoch lengths, and sampling decisions that determine your result rather than merely displaying it. The MNE tutorials teach the how without the why, and a filter you cannot explain is a finding you cannot defend at review. That is a research act, not instruction, so no diagnostic skips it.
Buy on sale only, never at list price, and check the current local price rather than trusting any figure written here. The certificate is worthless; you are buying the teaching.
3.4 The rest of his catalogue, and why it is out
| Course | Verdict |
|---|---|
| Complete neural signal processing and analysis: Zero to hero | Out by default, one exception below. The most subject-matched material in existence for this lane: neural time series, time-frequency analysis, synchronisation, permutation testing. It loses on three counts. It is MATLAB throughout, so every exercise needs translating, which removes the main advantage a paid course has over his free lectures. It is 45+ hours, roughly four times the course you are buying, arriving in blocks where the project already owns your attention. And most of its depth sits beyond a first bounded capstone |
| A deep understanding of deep learning | Out. Excellent and Python-native, but nothing in a first capstone requires deep learning, and Neuromatch's deep learning curriculum is free and open |
| Complete linear algebra | Out. 30+ hours where 3Blue1Brown's free three hours covers what you need |
| Statistics and machine learning courses | Out. You teach statistics |
The one exception. If Gate A produces a capstone whose primary method is time-frequency decomposition or connectivity analysis rather than band power and standard features, then Zero to hero stops being general depth and becomes the named method in your Block 6 learning contract. Buy it then, work only the sections your method needs, and implement in MNE-Python. Reading MATLAB is easy after Block 2; writing it is unnecessary.
04 Description before classification –
You cannot classify what you cannot describe. The sequence, from the atlas, is fixed and you follow it every time.
Use the description form in templates.md §11.7 every time. Filling it in for a normal recording is not wasted effort; it is how the form becomes automatic before anything unusual appears.
You are producing descriptions, not diagnoses. "Sharply contoured waveform, left temporal, in drowsiness, without after-going slow wave" is a description you can defend. "Left temporal epileptiform discharge" is a diagnosis you cannot.
05 Sequence across the blocks –
| Block | Status | Topic | Artifact |
|---|---|---|---|
| 1 | Core | Scope, measurement chain, safe-statement rewrites | EEG-SCOPE-CARD.md |
| 2 | Core | Acquisition, electrodes vs channels vs derivations, references, montages, phase reversal | Channel map plus hand-computed montage arithmetic |
| 3 | Core | Normal adult awake, drowsy, and sleep; age dependence | Three structured descriptions from named figures in the normal EEG chapter |
| 4 | Core | Artefacts | Five cards from named figures in the artifacts chapter for eye, muscle, pulse, electrode pop, sweat, plus a separate fixed 50/60 Hz mains example; intervals marked on three excerpts |
| 5 | Core | The two matched mimic contrasts (section 6) | Two comparison cards (AES Figs 48/62 and 49/73); the five-packet transfer test |
| 6 | Core | Label provenance and reference-standard reliability | Data cards for your corpora |
| Block 6, if detection project | Conditional | Remaining five variants | Remaining comparison cards |
| If EEG capstone | Conditional | Harmonisation, domain shift, preprocessing policy | Corpus label dictionary, transfer-risk register |
Egypt runs 50 Hz mains; most North American recordings, including CHB-MIT, are 60 Hz. A notch filter must match the source, not your location. Nothing warns you when it does not.
06 Benign variants and mimics: the actual procedure –
Why this matters scientifically, not just clinically. Benign variants are among the common sources of false positives for an automated detector. Knowing them is how you interpret your own error analysis rather than reporting a number you cannot explain. This is the clearest place in the whole roadmap where your medical background becomes a technical advantage.
6.1 The set, and what is core
Per the atlas the full set is: wicket waves, rhythmic mid-temporal theta of drowsiness (RMTD), benign small sharp spikes (BSSS), 14-and-6 positive spikes, 6-Hz phantom spike-and-wave, subclinical rhythmic EEG discharge of adults (SREDA), and midline theta.
Core, in Block 5: the two matched contrasts below. The card template has a column per side, so that is two cards, not seven. Work them from AES Figures 48 and 62, and 49 and 73.
Conditional, and narrower than it looks: the remaining five variants. They matter when your capstone involves epileptiform or seizure detection, because that is when a false positive is likely to be one of them. They are optional for an EEG project that is not detection-related, such as a sleep, ERP, connectivity, or artefact-robustness question.
If they apply, they attach to Block 6 as part of the project-learning contract and replace equivalent learning-contract work rather than adding to it, not to a block of their own. If they do not apply, say so once in writing and move on.
6.2 The card, thirteen fields
One card per variant. Every field filled from the source, not from memory or from this file.
| # | Field | What goes in it |
|---|---|---|
| 1 | Age group | Where it typically appears |
| 2 | State | Awake, drowsy, light sleep, deep sleep |
| 3 | Topography | Where on the scalp |
| 4 | Laterality | Unilateral, bilateral, shifting |
| 5 | Morphology | Shape, described in standard terms |
| 6 | Frequency | Hz, and whether it varies |
| 7 | Amplitude | Typical range in microvolts |
| 8 | Duration and rhythmicity | Brief train, sustained run, isolated |
| 9 | Evolution | Present or absent, and this is often decisive |
| 10 | After-going slow wave | Present or absent |
| 11 | Background disruption | Does surrounding activity change |
| 12 | Closest pathological mimic | The thing it is mistaken for |
| 13 | Discriminating constellation | The set of features that together separate them. Not one magic sign |
6.3 The level of detail required
Worked from the atlas, so you can see what a finished card looks like. Wicket waves are among the most commonly encountered benign variants and a recognised cause of EEG over-reading. They occur in brief trains or clusters, have an arciform appearance, are most frequent over temporal regions and may be unilateral or bilateral, run at roughly 6 to 11 Hz with amplitudes in the region of 60 to 200 microvolts, appear mainly in older adults during drowsiness and light sleep, and critically are not accompanied by an after-going slow wave.
That last clause is part of field 13, and only part. A card is finished when it records the constellation: evidence supporting the variant, evidence supporting the mimic, evidence against each, residual uncertainty, and the displayed montage, reference, filters, gain or sensitivity, and time scale where known. No single feature is decisive on its own.
6.4 The two contrasts that carry the module
Matched pairs, because a constellation only becomes memorable in contrast.
The constellation is the after-going slow wave together with background disruption, field, and state context. Sharpness alone decides nothing.
The constellation centres on evolution: RMTD is monotonous and state-dependent, while an ictal pattern evolves in frequency, amplitude, or morphology and may spread. Rhythmicity alone decides nothing.
Write both contrasts out in full before attempting the transfer pack. These two are the core requirement; the other variants wait for a deepen-EEG decision.
Why sharpness is not enough. Superficial sharpness or rhythmicity is present in benign variants, artefacts, and genuine discharges alike. Any rule you build on sharpness alone will overcall. The discriminators are contextual: state, evolution, after-going slow wave, field, and background disruption.
6.5 The transfer test
Five fixed textual packets in templates.md §11.11, with a concealed key. Each describes a recording in the same structured fields you have been using. You classify each as: normal physiological pattern, benign variant, artefact, possible epileptiform or ictal pattern, or insufficient information.
At least four of five broad categories correct, including both benign-versus-pathology discriminators, with zero categorical diagnostic claims.
On failure: two to four focused hours of remediation on the specific confusions, then one retest. A second failure means you stop making independent pattern claims. The computational work continues, and any interpretation-sensitive decision requires expert adjudication with that stated in your outputs.
Bank discipline. Once you have seen a key, that item is retired from testing. Keep practice and test material separate from the start.
What a larger bank can honestly be. Textual packets test reasoning, not visual pattern recognition. A genuine held-out waveform bank requires real recordings you have not seen, ideally with adjudicated labels, and without a supervisor you cannot construct one that proves visual competence. Say that plainly rather than implying otherwise. Textual transfer is what this route can honestly assess.
07 Label provenance, the scientific version –
Block 6 and beyond, and the part that most directly feeds your capstone. For every corpus you use, answer in writing:
- Who annotated it, and what was their expertise?
- Against what definition or criteria?
- Was any agreement checking done, and what did it show?
- What was the annotation unit: event, window, or record?
- What happens to your measured performance if those labels are imperfect?
Interrater agreement on EEG interpretation is imperfect and this is well documented in the literature. Find the actual reported figures for your corpus and read them properly rather than citing a secondhand number, including any number in this file.
Your model's measured ceiling is set partly by annotation reliability, not only by your method. That belongs in your limitations section, and understanding it is the difference between a competent paper and one that gets cited.
08 If your capstone is not EEG –
Keep sections 1 to 4, the two matched contrasts in section 6, and section 7. Drop the rest. Measurement literacy, description discipline, and label provenance transfer to every modality; the variant catalogue does not.
If Gate A sends you to population spikes, imaging, or behaviour, run the equivalent of section 7 on that modality instead: who made the labels, how, and with what reliability.