From noise to structured interference
In Vol I you learned to band-pass and notch. That is enough when the enemy is broadband thermal noise. But open a real EEG or ECoG file and the picture inverts: a single blink can be 100–200 µV where the cortical rhythm you want is 5–20 µV. The non-brain content does not sit politely in a separate band — an EMG burst is broadband, a motion transient overlaps every rhythm. The graduate reframing: an artifact is not 'noise' to be filtered but structured interference to be modelled and separated.
Define it operationally: an artifact is any recorded variance that does not originate in the neural process you are trying to measure. That definition is deliberately relative — a saccade is an artifact to a motor-imagery decoder but the signal of interest to an eye-movement study. What counts as artifact is set by your question, which is why there is no universal 'clean' recording, only a recording clean enough for a purpose.
A generative, linear-mixing view
Nearly every removal method — regression, SSP, ICA, ASR, beamforming — is a special case of one model. Volume conduction is, to good approximation, linear and instantaneous, so the channels-by-time data matrix \mathbf{X} is a mixture of source time courses \mathbf{S} through a mixing matrix \mathbf{A} set by geometry and conductivity.
\mathbf{X} = \mathbf{A}_{\text{brain}}\mathbf{S}_{\text{brain}} + \mathbf{A}_{\text{art}}\mathbf{S}_{\text{art}}Every channel is a weighted sum of neural and artifact sources; cleaning means estimating and removing the second term without disturbing the first.
What you record is a blend: every electrode picks up a weighted mix of the real brain signals plus junk like blinks, muscle, and heartbeat. Cleaning means pulling the two mixtures apart and discarding the artifact half without touching the brain half.
- \mathbf{X}
- The raw recording you actually measure (channels by time).
- \mathbf{S}_{\text{brain}},\ \mathbf{S}_{\text{art}}
- The hidden neural sources and the hidden artifact sources (eyes, muscle, heart) — what you wish you could see directly.
- \mathbf{A}_{\text{brain}},\ \mathbf{A}_{\text{art}}
- The mixing weights: how strongly each source spreads to each electrode.
An eye blink is one artifact source; it appears large on frontal channels and small at the back — that spatial spread is one column of \mathbf{A}_{\text{art}}.
This one equation organises the whole field. Methods differ only in how they get a handle on \mathbf{A}_{\text{art}} or \mathbf{S}_{\text{art}}: regression uses a measured reference (EOG/ECG); SSP and ICA estimate artifact topographies (columns of \mathbf{A}) from the data; ASR estimates a clean covariance and rejects directions that violate it. Keep the picture in mind and each method becomes a variation on a theme rather than a separate trick.
A working taxonomy
Sort artifacts by origin, because origin predicts the signature you exploit. Physiological: ocular (blinks, saccades; large, frontal, low-frequency), EMG (broadband >20 Hz, temporalis/neck), cardiac/pulse (~1 Hz, ECG-locked), and motion (low-frequency, mechanically coupled). Environmental: mains line noise at 50/60 Hz and harmonics, plus electromagnetic interference. Instrumental: electrode pops, drift and DC offset, and bad channels. Interface-specific: for closed-loop and bidirectional systems, stimulation artifacts that dwarf the neural signal.
Signal quality is a number, not a vibe
Because artifacts are quantifiable, so is what survives. The workhorse is the signal-to-noise ratio and its most important lever, coherent averaging: repeat an event N times and average, and the time-locked signal reinforces while noise partly cancels.
\mathrm{SNR}_{\bar{x}} = \sqrt{N}\;\mathrm{SNR}_{x}Averaging N repeats improves amplitude SNR by √N — the single most-used quality lever in evoked BCI.
Repeat the same stimulus many times and average the responses: the random noise partly cancels while the real evoked signal stays put, so quality grows with the square root of the number of repeats.
- N
- Number of repeated trials you average together.
- \mathrm{SNR}_{x}
- Signal-to-noise ratio of a single trial.
- \mathrm{SNR}_{\bar{x}}
- Signal-to-noise ratio after averaging.
- \sqrt{N}
- The gain factor — the payoff for averaging.
Average 100 trials and SNR improves 10-fold; to double it again you must gather 4 times more trials.
Why decoders raise the stakes
In a classical ERP study an artifact mostly adds variance and costs you statistical power. In a BCI it can do something worse: if the artifact correlates with the task, the features carry it and the decoder will happily learn it. A 'brain' speller that in fact reads frontal EMG or eye position can post excellent accuracy that collapses the moment the user stops tensing or looking. This is leakage wearing a lab coat, and it is why artifact handling is not hygiene you do at the end — it is part of the claim you are making.