Page 169 - Handbook of Biomechatronics

166 Domen Novak

two key concepts for future BCI research: BCI training should be optimized
for a particular application, and new BCI users should be provided with
advice on how to effectively make use of a BCI in a particular application
(e.g., how early to perform motor imagery in order to compensate for
delays). To the best of the author’s knowledge, little systematic research
has been done in this direction, and would represent a promising topic
for future work.
Furthermore, as emphasized by studies of ERPs for error correction and
error-driven learning, the BCI should also adapt to the user. Several strat-
egies for such ERP-driven adaptation have now been proposed and vali-
dated, but have largely been limited to adapting the BCI itself.
A promising direction that is still in its infancy would be to use ERPs to adapt
the behavior of other machines, as demonstrated by Chavarriaga and Milla ´n
(2010). This is a much greater challenge than adapting BCIs since it is often
unclear how to adapt a machine in response to a detected ERP, for example,
we may not be able to determine what specific action caused the ERP or
what a more appropriate action would be in that specific situation. None-
theless, addressing this challenge would greatly broaden the impact of BCIs
by creating a new generation of intelligent biomechatronic devices that are
responsive to the users’ mistakes, preferences, and dislikes.

3.4 Comparison to Other Control Methods

Finally, if BCIs are to achieve widespread adoption, their potential benefits
must be made clear to users. While many studies have demonstrated strong
benefits of BCIs in applications such as communication, some areas still suf-
fer from unclear usefulness of the technology. One such area is the use of
passive BCIs for estimation of mental workload and consequent task diffi-
culty adaptation: while many studies have demonstrated that EEG-based dif-
ficulty adaptation achieves better results than performance-based adaptation,
it is unclear whether the improvement is sufficient to justify the additional
cost, setup time, and inconvenience for the user. This issue has been empha-
sized by recent reviews of passive BCIs (Brouwer et al., 2015), and is doubly
complicated since many studies report only the classification accuracy (e.g.,
ability to discriminate between high and low workload) of an EEG-based
method compared to a performance-based method instead of reporting
the effect on the user’s enjoyment, learning rate, or other important out-
come. The classification accuracy, particularly when calculated offline on
prerecorded data, does not necessarily have a clear relationship to BCI