Non-invasive information rate is far below ordinary input

What it is

A brain–computer interface (BCI) recorded from the scalp typically moves a few to a few tens of bits per minute on information transfer rate (ITR). Mice, keyboards, and touch sit orders of magnitude above that. The non-invasive channel is not a slightly slower keyboard; it is what remains after skull filtering and volume conduction have thinned the bandwidth.

Why it happens

Scalp EEG mixes cortical postsynaptic potentials after they travel through cerebrospinal fluid, skull, and scalp. The skull is a low-pass filter; spatial resolution is centimetre-scale, and many sources are inseparable at the scalp. Inferring cortical sources from scalp potentials is an ill-posed inverse problem: solutions are non-unique and noise is amplified. Usable features therefore collapse to a few coarse, stable dimensions—event-related desynchronization of sensorimotor rhythms, time-locked potentials such as P300, frequency tags in steady-state visually evoked potentials. Each choice still needs several trials to lift SNR over a classifier threshold, so trial structure spends the clock. The ITR formula multiplies accuracy, class count, and time per trial; any one of those going sour drops the rate.

Studying it

Spellers, two-dimensional cursor control, and binary switches are the classic tasks. Factors include stimulus rate, number of classes, and whether trials are averaged. Outcomes should report accuracy, bits per minute, and correct selections per minute together; accuracy alone hides “slow and precise.” The distribution of ITR across people and sessions is more informative than a peak. When comparing a BCI with a keyboard, both should finish the same text or selection task with calibration time included. Treating a well-trained laboratory SSVEP peak as product bandwidth overstates by about an order of magnitude.

Where it stops holding

Invasive arrays have higher bandwidth; they do not underwrite a scalp headset. SSVEP under strong flicker can approach a hundred bits per minute, at the cost of visual load and photosensitivity risk, and that rate does not describe quiet imagery. Residual muscle artifact, mains interference, and a lifted electrode can drive field ITR to near zero. Rate also moves with state: drowsiness drifts rhythmic features. Drawing non-invasive BCI as “typing with thought” skips the physical cap of skull filtering.

Applying it

• Plan tasks by correct selections per minute, not by “it can classify”: a few discrete commands with long confirmation, not a continuous pointer.
• Quote the rate that remains stable after calibration, not the laboratory peak.
• Do not leave an action only on EEG if a key, gaze confirm, or switch can do it.
• Verify by measuring ITR and completion time on the target users and electrode kit across several days, against that user’s existing assistive path.

Related

• Same group: C9.02.2 Training cost and individual differences are the main barriers · C9.02.3 Currently suited as a replacement channel rather than an enhancement
• Adjacent: C9.07 Capability Bounds of BCIs · C9.11 Latency of Physiological Input
• Search: information transfer rate · volume conduction · non-invasive EEG