Content uploaded by David Steyrl
Author content
All content in this area was uploaded by David Steyrl on Oct 12, 2016
Content may be subject to copyright.
G
R
A
Z
B
C
I
R
A
C
I
N
G
T
E
A
M
BCI adaptation for end users
The Graz-BCI approach
Andreas Schwarz1, D. Steyrl1, M. K. H¨oller1, K. Statthaler1and G.R. M¨uller-Putz1
andreas.schwarz@tugraz.at, gernot.mueller@tugraz.at
1Institute of Neural Engineering, Graz University of Technology, Austria
Non-invasive Brain-Computer interfaces (BCI) enable its users to interact with their environment only by thought. A possible BCI
application may be to control a computer game by e.g. imagery of motor tasks. However, this requires several control commands and individual
BCI training. In the following, we describe our four stage approach for individualizing and adapting BCI technology for an end user based on the
performance of the pilot of the MIRAGE91 racing team [1].
Our approach is based on [2] and the findings of Friedrich et al.[3](Figure 1).
Pre-Screening
basic Motor Imageries
& rest Condition
BCI capability &
user compliance
Screening
Proceed?
7 Mental strategies
& rest Condition
Find best combination
performance & compliance
BCI +Feedback
Best 4-class
combination
Game based training
Best 4-class
combination
Evaluate Tune
Test feedback &
feedback compliance
III III IV
REPEAT
Figure 1: 4-Stage training procedure. Stage I to III investigate the best performing mental tasks for the user, which are applied in Stage IV for game based training.
In Stage I (Figure 2) we performed pre-screening to test users’ BCI capability and compliance. Results of this stage indicated whether
continued training with the user was reasonable. Stage II (Figure 3) incorporated a screening of several mental tasks, including a non-control
state. In an offline cross-validation procedure, we determined the most effective (in terms of accuracy and user acceptance) combination of at least
4 different classes. In stage III (figure 4), the previously identified class combination was used to test user compliance to feedback. In the
beginning of Stage IV (figure 5), a BCI was closely tailored to users based on the findings in the previous stages. Thereafter the user started
game based BCI training.
Figure 2: Cross-validation (10x5) results of Pre-Screening: Accuracy
over all trials and trial based confusion matrix. Standard GRAZ-BCI paradigm
was used to record 50 Trials per condition (TPC).
Figure 3: Cross-validation (10x5) results of Screening Accuracy over
all trials and trial-based confusion matrix of the best performing mental task com-
bination (out of 70).
-2 -1 012 3 4 5
10
20
30
40
50
60
70
80
90
100 Average
Figure 4: Stage III Online Performance: Accuracy over all trials and trial-
based confusion matrix. 50 TPC were used for training the BCI model. Therafter
40 TPC were recorded where the user received feedback for evaluation.
10 trials per condition - runtime (s)
2| 6
2 | 6
2 | 6
3 | 6
0 | 7
0 | 7
1 | 7
1 | 7
100
120
140
160
180
200
Sessions 9 - 18
Figure 5: Stage IV game based training over several sessions. The
median race time over sessions 9 to 18 shows a steady decrease.
References
1. A. Schwarz, D. Steyrl et al. ”Brain-Computer Interface adaptation for an end user to compete in the Cybathlon”, IEEE International Conference on
Systems, Man, and Cybernetics (SMC 2016), At Budapest, Hungary, 2016, accepted
2. G. M¨uller-Putz, R. Scherer, et al., “Temporal coding of brain patterns for direct limb control in humans,” Frontiers in Neuroscience, vol. 4, no. 34,
2010.
3. E. Friedrich, C. Neuper, et al., “Whatever works: A systematic user-centered training protocol to optimize brain-computer interfacing individually,”
PLOS ONE, vol. 8, no. 9, 2013.
Acknowledgments
