Reading raw traces

Why this matters

Every analysis you will ever run starts with someone deciding whether the file is worth analysing, and that decision is made by looking at the traces. A viewer set to the wrong scale hides a bad channel; a window that is too long turns alpha into a smear; a montage chosen without thought puts the reference channel’s problems on every trace. This lesson is the skill the rest of the ladder assumes: open a raw file, set it up so that it can be read, and read it.

Concepts

The viewer is an instrument: sensitivity, window, montage

A raw EEG viewer has three settings that decide what you can see, and none of them has a “correct” value.

Sensitivity (µV/cm, or µV per division). Clinical EEG is traditionally displayed at a fixed sensitivity so that amplitude can be judged by eye against the page; a common convention is 7 µV/mm (TODO(confirm)). In a scripted viewer such as mne.io.Raw.plot() the equivalent control is scalings: the number of µV that fills one channel’s slot. Set it so that the largest ordinary activity (posterior alpha in a relaxed subject, roughly tens of µV peak-to-peak) fills about half the slot. Too small and blinks overlap five neighbours; too large and a 30 µV alpha train becomes a flat line and a flat channel becomes invisible.

Window length. Ten seconds per page is the clinical convention and a good default: alpha at around 10 Hz shows about a hundred cycles, individual blinks and pops are resolvable, and slow drift is still visible as a tilt. Use 1–2 s when you need to see waveform morphology (mu arches, spindle shape), and 30–60 s when you are looking for state changes and slow drift. The same file at three window lengths is three different pictures; learn to switch.

Montage. The traces you see are differences between electrodes. In a referential display each trace is one electrode minus the recording reference (or minus a re-computed reference such as the average). In a bipolar display each trace is the difference between two neighbouring electrodes along a chain, which cancels widespread activity and emphasises local gradients. Referential displays show you the reference’s problems everywhere (a noisy reference electrode makes every channel look noisy in the same way); bipolar displays hide them but also hide widespread activity. For a first look, referential to the recording reference, then average reference, then bipolar if something looks odd — and always know which one you are looking at.

Channel ordering. Order channels anatomically (left frontal to occipital, then right, or interleaved by row) rather than by acquisition index. Lateralized and regional features are only obvious when neighbours are adjacent on the screen.

Awake rhythms

The awake, relaxed adult EEG has three rhythms you must be able to point to on a screen.

Posterior alpha. A roughly sinusoidal rhythm in the conventional 8–13 Hz band, largest over occipital and parietal electrodes (O1, O2, P3, P4, Pz), waxing and waning in spindle-like trains over a few seconds. Its amplitude varies a lot between people; its defining property is reactivity: it appears when the eyes close and is attenuated by eye opening, visual attention and mental effort. The frequency of a person’s own alpha peak is their individual alpha frequency, which you will measure in L1.7.

Mu. A central rhythm (C3, C4) in a similar frequency range to alpha, but with a characteristic arch- or comb-shaped waveform (sharp on one side, rounded on the other) rather than a sinusoid. It is not blocked by opening the eyes; it is attenuated by movement, by preparing to move, and by imagining movement of the contralateral limb. Alpha and mu overlap in frequency, so a spectrum cannot tell them apart — topography, waveform and reactivity can.

Beta. Low-amplitude, faster activity (conventionally 13–30 Hz), most visible over frontal and central sites, usually riding on top of slower activity. Some beta is normal; a sudden increase in fast, spiky activity that follows jaw or neck tension is muscle, which L0.5 covers.

State changes

Alpha reactivity (eyes open vs eyes closed). When the eyes close, posterior alpha rises within about a second; when they open, it drops. On a 30-s window the transition is visible as a change in the texture of the posterior channels: from low-amplitude, irregular activity to regular, higher-amplitude trains. This is the most reliable single feature for telling the two states apart by eye, and it is the basis of this lesson’s exercise. Note that a small fraction of healthy people have little visible alpha at any time (TODO(confirm) proportion); for them the eyes-closed state is harder to call from the trace alone.

Drowsiness. As a subject drifts toward sleep the posterior alpha fragments and drops out, activity slows toward theta, and slow rolling (roving) eye movements appear as low-frequency deflections in the frontal channels. Vertex sharp waves at Cz may appear as sleep onset approaches. Drowsiness is common in long resting recordings and in the second half of any boring task; a “resting-state” file with drowsy segments is a mixture of two states, not one.

Sleep spindles. Brief trains of roughly 11–16 Hz activity lasting about 0.5–2 s, most prominent over central and frontal sites, that mark stage N2 sleep. A subject who was supposed to be awake and shows spindles has fallen asleep; the file’s “resting” label is wrong for those minutes.

The data behind this lesson

• ds-eegbci: 64-channel, 160 Hz, no hardware filters. Run R01 is one minute eyes-open, R02 one minute eyes-closed, recorded as separate files — clean examples of each state, but no transition inside one recording.
• ds-lemon: 62-channel, raw release at 2500 Hz, sixteen alternating one-minute eyes-closed and eyes-open blocks in one recording. It is the only spine source of a real within-recording transition, which is what the exercise needs. The widget serves a resampled 250 Hz segment around a block boundary.
• ds-dortmund (fallback if ds-lemon is dropped, §13 items 5 and 17): 3-minute eyes-closed and 3-minute eyes-open blocks recorded as separate files, resampled to 250 Hz for the widget. Long single-condition segments, but no within-recording transition.

All widget segments are derivatives (cropped, resampled, channel-subset) of these public datasets; the provenance line under the widget names the dataset, subject, run and license.

Explore

Work through the segments in this order: ds-eegbci R02 (eyes closed) at 10 s per page; the same at 2 s per page to see alpha waveform; ds-eegbci R01 (eyes open) at 10 s; then the multi-minute segment at 30 s per page with annotations off. Turn annotations on only after you have made your own call.

Annotations in this widget are label_source: algorithmic until the author reviews them (§4.5) — treat them as a second opinion, not ground truth.

Practice

The notebook opens the same files in the MNE raw browser, sets scalings, duration and n_channels, switches between the recording reference and an average reference, and adds and edits annotations you can save back to the file.

Exercises