Signal Noise And What A Brain Recording Contains

Raw data from any brain recording is dominated by things that are not brain activity. What counts as signal and what counts as noise depends entirely on the question being asked.

The body generates competing electrical activity

Muscles produce electrical potentials that are far larger than those from neurons, and the muscles of the jaw, neck and eyes sit close to any scalp recording.

Blinking, swallowing or clenching creates artefacts that can swamp the underlying activity for a period considerably longer than the movement itself.

Because these events correlate with what a person is doing, they are especially dangerous. An artefact that appears whenever a task begins can be mistaken for a task-related brain response.

The environment contributes its own interference

Mains electricity radiates at a fixed frequency, and any recording made indoors picks it up. Filtering it is routine but removes a slice of genuine signal alongside it.

Nearby equipment, moving cables and even the amplifier itself add further contributions, some of which vary with temperature or humidity across a session.

Careful grounding, shielding and short cable runs reduce these problems more effectively than any processing applied afterwards.

Scale determines what can be seen

A scalp electrode averages the activity of enormous numbers of neurons, blurred by the skull. It captures rhythms and broad states rather than individual cells.

An implanted array can resolve the firing of single neurons, but it sees only the small population near the electrode and nothing about the rest of the brain.

Neither is a complete picture, and claims about decoding capability mean very different things depending on which scale produced the data.

Processing choices shape the result

Filtering, re-referencing and artefact rejection all involve decisions, and different reasonable choices can produce visibly different conclusions from identical raw recordings.

Aggressive cleaning removes noise but can also remove real activity, particularly slow components that overlap with drift and movement in frequency.

Good practice is to report the processing steps in enough detail that another group could reproduce them, because the pipeline is part of the finding.

Averaging hides variability that matters

Many neural effects are only visible after averaging across many repetitions, because single trials are too noisy to interpret individually.

That averaging discards trial-to-trial variation, which is precisely what a real-time interface must contend with when it has one attempt at each decision.

This gap explains why an effect that is statistically solid in an offline analysis may still support no usable device at all.