Machine learning-based co-adaptive calibration: A perspective to ...

Machine learning-based co-adaptive calibration: A perspective to alleviate BCI illiteracy

Carmen Vidaurre,

Claudia Sannelli, Klaus-Robert Müller, Benjamin Blankertz

Machine Learning Laboratory, Berlin Institute of Technology

corresponding author:

[email protected]

Introduction: Classic BCI paradigm

A prior study (Berlin + Tübingen) performed the screening of 80 inexperienced users in calibration ⇒ feedback session (classic machine learning BCI paradigm). Calibration R

R

R

L

L

L

R

L

Feature extraction Training of classifier

Feedback

apply online in sliding windows

Motivation

•

The

BCI Illiteracy Phenomenon

20-30% of BCI users do not reach the level needed for control ∼ 70% binary classication [Kübler et al. 2004].

•

•

User categorization

Results of a previous study with N =80 naïve users and classical system design with calibration and feedback (Berlin+Tübingen):  Category I: Good calibration and good feedback ∼ 61%  Category II: Good calibration and bad feedback ∼ 14%  Category III: Bad calibration, no feedback possible ∼ 25% Machine Learning Co-Adaptive Learning [Vidaurre and Blankertz, 2009] System design that helps BCI-users to achieve good feedback.

Experimental setup

14 volunteers: 4 novice (no categorization) and 10 CO-ADAPTIVE CALIBRATION

Cat. III

FEEDBACK APPLICATION

DATA BASE

LAP C3, Cz, C4 BANDS [8-15] and [15-35] Hz

1

Sel. LAP SUBJECT SPECIFIC BAND 2

3

CSP SUBJECT SPECIFIC BAND 4

5

 Feedback immediately [Vidaurre et al. 2006, 2007]  3 dierent levels of adaptation [Vidaurre and Blankertz 2009]  First run 40 trials/class. Subsequent runs 50 trials/class

6

Level 1: 1 run

DATA BASE

EEG

LAP C3, CZ, C4

BCI-Paradigm

BAND PASS 8-15 Hz, 15-35 Hz

LDA

Feedback

ADAPT SUPERVISED Σ -1 , µ , µ 1

2

40 trials/class LR LF FR

l Pre-trained classier. Simple methods: fast adaptation l Lap. C3, Cz y C4, log band-power (8-15 and 15-35 Hz) l Supervised adaptation: class means and cov. matrix [Vidaurre et al. 2006,2007] l 3-classes in 3 binary groups (LR, LF, FR), 40 trials/class

Level 2: 2 runs

Run 1 2-classes selected

EEG

6 SELECTED LAPLACIANS

RESELECT SUPERVISED

SUBJECT SPECIFIC BAND PASS

Regularized LDA

Feedback

RETRAIN SUPERVISED

l Classes, band, Lap. channels from run 1 l 6 sel. laplacians updated after every trial l Supervised adaptation: retrain classier [Vidaurre and Blankertz, 2009]

Level 3: 3 runs

Runs 2-3

EEG

CSF

SUBJECT SPECIFIC BAND PASS

LDA

Feedback

ADAPT UNSUPERVISED µ

l Band and CSF xed from run 2-3 l Unsupervised adaptation by bias estimation (real performance) [Vidaurre et al. 2008]

Performance by Category

BCI feedback accuracy [%]

100 90 80 70 60

1 Level 1

2−3 Level 2

4−6 Level 3

50 40

Novice (N=4)

BCI−Control (N=5)

No BCI−Control (N=5)

 Performance increase with more complex ML methods (run 2)  All 4 novice users achieved good performance  5 Cat. III users overcame BCI-illiteracy  2 Cat. III users almost reached the level criterion (∼67%)  3 Cat. III did not improve over the session

Analysis of one Cat. III user

sgn r2

sgn r2( foot run 1 , right run 1 )

Cz

CP3

Cz

CP3

0.1

0.05

0

sgn r2( foot run 6 , right run 6 ) −0.05

−0.1

Discussion

• The simple methods of Level 1 allowed rapid adaptation and good performance for novice users. • Machine Learning methods (levels 2 and 3) helped 5 Cat. III users to achieve control. • 2 users more were at the limit of the control threshold. • Unfortunately, 3 Cat. III participants did not reach control ⇒ Operant Conditioning? • The binary grouping of classes is not consistent (block-eect).

Thanks for your attention References

¬ Kübler, Neumann, Wilhelm, Hinterberger, Birbaumer, Predictability of Brain-Computer Communication, J. Psychophysiol., 2004 ­ Vidaurre and Blankertz, Brain Topography, 2009 ® Vidaurre, Schlögl, Blankertz, Kawanabe, Müller, Unsupervised adaptation of the LDA classier for Brain-Computer Interfaces Proc. 4th Int BCI Workshop and Training Course 2008, Verlag der TUGraz, 2008 ¯ Vidaurre, Schlögl, Cabeza, Scherer, Pfurtscheller, Study of On-Line Adaptive Discriminant Analysis for EEG-Based Brain Computer Interfaces, IEEE tbme, 2007 ° Vidaurre, Schlögl, Cabeza, Scherer, Pfurtscheller, A fully on-line adaptive BCI, IEEE tbme, 2006