Page 161 - Handbook of Biomechatronics

158 Domen Novak

from the exercise; however, if the exercise is very difficult, the patient will
become annoyed, lose focus, and not wish to continue. By monitoring the
patient’s workload level and using it to adapt the exercise difficulty, the BCI-
controlled system can achieve optimal rehabilitation outcome. Admittedly, a
similar difficulty adaptation could be achieved in a much simpler way by
simply monitoring the patient’s task success rate and using it as a basis for
adaptation. However, this would not capture the patient’s internal mental
state and would potentially be less reliable, for example, if a patient has a
low success rate, it is possible that they are overwhelmed by the task and need
an easier one, but it is also possible that they are bored by the task and not
putting any effort into it, or that they are trying hard and failing but still
enjoying themselves.
Estimation of patient workload for purposes of exercise adaptation in
motor rehabilitation was proposed as early as 2007 (Cameira ˜o et al.,
2007), and was first implemented using autonomic nervous system responses
as workload indicators (Novak et al., 2011), but EEG as a workload indicator
was implemented soon afterwards (Novak et al., 2015; George et al., 2012;
Park et al., 2015). The closed-loop approach is largely independent of the
type of physiological measurement: a rehabilitation robot adapts either its
level of assistance or the difficulty of the overall task (e.g., required speed,
range of motion) based on the inferred workload. An example a BCI-
controlled rehabilitation robot is shown in Fig. 6. However, the main weak-
ness of this technology is its unclear benefit: while some studies have shown
that, for example, physiology-based exercise adaptation is more accurate
compared to a “ground truth” than simple task-success-based adaptation
(Novak et al., 2011), there is so far no evidence that physiology-based adap-
tation results in better rehabilitation outcome. Thus, adoption of BCI-based
adaptation in clinical rehabilitation practice is unlikely until its benefits are
more clearly demonstrated.
Aside from motor rehabilitation, several other learning environments
could benefit from BCI-based difficulty adaptation. For example, Walter
et al. recently developed arithmetic learning software that automatically
adapts the difficulty of the presented material based on the learner’s EEG
(Walter et al., 2017). The EEG-based software was compared to a version
that only adapted the difficulty of the material based on the learner’s success
rate, and the EEG-based version was found to result in a higher learning
effect, though the difference was not statistically significant. This presents
the same challenge as BCIs for adaptation of rehabilitation difficulty: while
the EEG-based system appears to have short-term advantages over a purely