Biomedical Engineering Reference
In-Depth Information
Table 8.2
Classification Accuracies of Three Phases
Time
Window
(seconds)
Phase 1
Accuracy (%)
Phase 2
Accuracy (%)
Phase 3
Accuracy (%)
Subjects
Passband (Hz)
S1
10-35
2.5-8
94.00
98.11
97.03
S2
13-15
2.5-7.5
94.67
97.56
95.74
S3
9-15
2.5-7
74.71
80.13
81.32
S4
10-28
2.5-6
68.00
77.00
68.40
S5
10-15
2.5-7.5
66.00
72.22
71.50
Mean
—
—
79.48
85.00
82.80
tively. For subjects S1 and S2, no significant difference existed between the classifi-
cation results of the three binary classifiers, and a high accuracy was obtained for
three-class classification. For the other three subjects, the foot task was difficult to
recognize, and the three-class accuracy was much lower than the accuracy of classi-
fying left- and right-hand movements. This result may be caused by less training of
the foot imagination, because all of the subjects did more training sessions of hand
movement in previous studies of two-class motor imagery classification [55]. The
average offline accuracy was about 5% higher than the online training phase due to
the employment of parameter optimization and the CSP algorithm applied to
multichannel EEG data.
8.3.3.4 Phase 3: Online Control of Three-Direction Movement
In phase 3, a similar online control paradigm as in phase 1 was first employed to test
the effect of parameter optimization, and a 3% increase in online accuracy was
observed. Then, three of the subjects participated in online control of three-direc-
tion movement of robot dogs (SONY, Aibo) for mimicking a brain signal controlled
robo-cup game, in which one subject controlled the
goalkeeper
and the other con-
trolled the
shooter
. This paradigm and approach could be used for applications
such as wheelchair control [57] and virtual reality gaming [58, 59].
8.3.4 Alternative Approaches and Related Issues
8.3.4.1 Coadaptation in SMR-Based BCI
As discussed in Section 8.1.2, the BCI is not just a feedforward translation of brain
signals into control commands; rather, it is about the bidirectional adaptation
between the human brain and a computer algorithm [2, 6, 60], in which real-time
feedback plays a crucial role during coadaptation.
For an SSVEP-based BCI system, the amplitude modulation of target EEG sig-
nals is automatically achieved by voluntary direction of the gaze direction and only
the primary visual area is involved in the process. In contrast, for an SMR-based
BCI system, the amplitude of the mu and/or beta rhythm is modulated by the sub-
ject's voluntary manipulation of his or her brain activity over the sensorimotor area,
in which secondary, even high-level, brain areas are possibly involved. Thus, the
Search WWH ::
Custom Search