Training cost and individual differences are the main barriers

What it is

The main barrier for non-invasive BCIs is often not “is there EEG,” but that some people never learn stable control and everyone else must recalibrate on each donning. The literature calls the first group BCI illiteracy (also BCI inefficiency): under a standard paradigm, accuracy sits near chance for the long run. Training time and between-person spread decide whether the channel can ship, more than peak bitrate.

Why it happens

Scalp features track anatomy: cortical folding, skull thickness, and whether a sensorimotor rhythm is even visible at rest all differ across people. Spatial filters such as common spatial patterns are estimated from calibration data; a shifted electrode or a changed impedance misaligns the filter with today’s sources. Motor imagery also asks the user to hold a strategy that is hard to inspect; when the strategy drifts, the features drift. Non-stationarity across sessions—drowsiness, caffeine, a crooked cap—makes yesterday’s classifier fail today. Cost therefore splits: some people never pass the calibration gate; those who do still spend the first tens of minutes of a session on calibration trials.

Studying it

“How many participants reached operable accuracy” matters more than the mean among successes. Factors: number of calibration trials, whether retraining happens across days, whether feedback is online. Outcomes: the fraction of people who clear an accuracy threshold, calibration duration, and the drop across sessions. Dropping people who never clear the threshold from the paper hides the barrier. Whether transfer learning or unsupervised adaptation truly cuts calibration should be tested on a zero-calibration first session, not only on fine-tuning in the second half of the same day.

Where it stops holding

Evoked potentials (P300, SSVEP) usually transfer across people more easily than motor imagery because the stimulus time-locks the response; they move the burden onto visual load and fatigue rather than cancelling individual differences. Invasive systems have their own surgical and upkeep costs and cannot argue that a scalp headset “just needs training.” Calibration curves from healthy undergraduates do not predict users with damaged motor cortex or fragile attention. Illiteracy rates rise when no experimenter is present to seat electrodes.

Applying it

• Put an abandonable calibration gate in onboarding: if accuracy stays near chance after a budgeted set of trials, route to another input rather than empty practice.
• Separate short within-session calibration from cross-day retraining; a product cannot assume users will sit a full paradigm every morning.
• Show personal stability, not a group-mean accuracy as the selling point.
• Verify without screening out “bad signal” recruits; report the distribution of time from power-on to first reliable selection, including failures.

Related

• Same group: C9.02.1 Non-invasive information rate is far below ordinary input · C9.02.3 Currently suited as a replacement channel rather than an enhancement
• Adjacent: C9.10 Individual Calibration and Baseline Drift · C9.07 Capability Bounds of BCIs
• Search: BCI illiteracy · session-to-session variability · CSP calibration