Page 164 - Handbook of Biomechatronics

Biomechatronic Applications of Brain-Computer Interfaces 161

but does not learn anything from the new data. Thus, even if operating con-
ditions change or the BCI keeps making mistakes, it will not change its pre-
viously programmed pattern-recognition algorithms. This puts the onus on
the user to learn how to use the BCI effectively, often by trial and error.
BCIs that incorporate ERPs, on the other hand, are able to detect that an
error has occurred and then take corrective actions. The ERP can be caused
either by an error on the part of the user or on the part of the machine, and
some studies (though not all) have indicated that larger errors evoke larger
ERPs (Gentsch et al., 2009; Sp€uler and Niethammer, 2015). An excellent,
detailed review of ERPs in BCIs is provided by Chavarriaga et al. (2014),
and we briefly summarize key developments in this section.

2.9.1 Error Correction
In a first report on the use of ERPs with BCIs, Schalk et al. (2000) demon-
strated that, when controlling a cursor with an EEG-based BCI, erroneous
control results in an ERP. Since then, several studies of ERPs in response to
successful and unsuccessful BCI use have shown that ERPs are relatively sta-
ble and occur reliably, allowing BCIs to determine whether the correct
desired command was selected based on the user’s EEG. Furthermore,
the amplitude and waveform of ERPs do not differ significantly between
tasks, suggesting that ERP analysis could be independent of the BCI type
and the biomechatronic device that it is controlling (Iturrate et al., 2011).
One of the earliest BCIs that used ERPs to correct errors was presented
by Milla ´n and Ferrez (2008), who used motor imagery to control a cursor.
After each cursor movement, the EEG was checked for ERPs that would
indicate an erroneous motion; if one was detected, the cursor was automat-
ically moved back to the previous position. Based on ERP detection, 80% of
motions were correctly classified as correct or erroneous, resulting in signif-
icantly improved cursor control. An interesting similar concept was pres-
ented by Artusi et al. (2011) with a simulated motor-imagery-based BCI:
the BCI analyzed the EEG and classified the type of motor imagery, then
showed the classification result to the user on the screen before acting on
it. If the user exhibited an ERP in response, the classification result was con-
sidered erroneous and discarded and the task had to be repeated. Both of
these studies showed high potential for ERPs to automatically identify
and correct erroneous BCI behavior, though they were conducted with
proof-of-concept rather than realistic BCI systems.
Following early proof-of-concept studies, ERP detection was widely
implemented in P300 spellers. Essentially, the P300 is used to select a