Closed-loop control (BCI as feedback control)
Closed-loop control refers to operating a BCI while its decoded output is fed back to the user in real time — usually as a moving cursor, a limb, or a speller cursor — so that the user can perceive the effector and continuously correct it. The defining consequence is that the neural signals are no longer statistically the same as those recorded during open-loop calibration: once feedback is present the user becomes an active controller, adjusting strategy trial by trial, so the input distribution the decoder must handle is coupled to the decoder itself. This coupling is why a BCI is a control system, not merely a pattern classifier.
The central empirical fact is that offline (open-loop) accuracy is a poor and sometimes misleading predictor of closed-loop performance. A decoder that scores highest on held-out calibration data can be unusable online because it responds to features the user cannot volitionally stabilize under feedback, while a decoder with mediocre offline numbers may be excellent online because the user learns to drive it. Evaluations that stop at cross-validated offline accuracy therefore routinely overstate or misrank real usability.
In cursor BCIs it is common to see a decoder that is 90 percent accurate offline yet produces a cursor that drifts and cannot hold a target, because the user has no way to counteract a bias that offline scoring never penalized.
Why closed-loop testing is not optional.
The single most repeated methodological error in BCI is treating offline accuracy as the endpoint. It is a screening tool at best; the only faithful measure of a control interface is control performance with the human in the loop.