23
1 Takumi Kodama , Kensuke Shimizu , Shoji Makino and Tomasz M. Rutkowski Full–body Tactile P300–based Brain–computer Interface Accuracy Refinement @bioSMART conference 2016 1 1 1 2, 3, 4 Life Science Center of TARA, University of Tsukuba, The University of Tokyo, Saitama Institute of Technology, RIKEN Brain Science Institute

Full–body Tactile P300–based Brain–computer Interface Accuracy Refinement

Embed Size (px)

Citation preview

1

Takumi Kodama , Kensuke Shimizu , Shoji Makino and Tomasz M. Rutkowski

Full–body Tactile P300–based

Brain–computer Interface Accuracy Refinement

@bioSMART conference 2016

1 1

1 2, 3, 4

1 23 4

Life Science Center of TARA, University of Tsukuba, The University of Tokyo, Saitama Institute of Technology, RIKEN Brain Science Institute

1: Introduction - What’s BCI?

● Brain Computer Interface (BCI)○ Neurotechnology ○ Exploits user intention ONLY using brainwaves

2

1: Introduction - ALS Patients

● Amyotrophic lateral sclerosis (ALS) patients○ Have difficulty to move their muscle by themselves○ BCI could be a communicating tool for them

3

ʳ!…

● Tactile (Touch-based) P300-based BCI paradigm○ P300 responses were evoked by external (tactile) stimuli○ Predict user’s intentions by decoding P300 responses

1: Introduction - Research Approach

41. Stimulate touch sensories 2. Classify brain response

AB

A

B

3. Predict user intention

92.0% 43.3%

A B

TargetNon-Target

P300 brainwave response

● Full-body Tactile P300-based BCI (fbBCI) [1]○ Applies six vibrotactile stimulus patterns to user’s back○ User can use fbBCI while lying down and interacting

using a whole body surface

1: Introduction - Our Method

5[1] Kodama T, Shimizu K, Rutkowski TM. Full Body Spatial Tactile BCI for Direct Brain-robot Control. In: Proceedings of the Sixth International Brain-Computer Interface Meeting. Asilomar Conference Center, Pacific Grove, CA USA: Verlag der Technischen Universitaet Graz; 2016. p. 68.

1: Introduction - Demonstration

6https://www.youtube.com/watch?v=sn6OEBBKsPQ

● P300 responses were confirmed (> 4 μV) in each channel

1: Introduction - fbBCI results (1)

7

TargetNon-Target

● Problem: Low online classification accuracies○ SWLDA : 53.67 % (10 users average)

1: Introduction - fbBCI results (2)

8

● To improve the fbBCI classification accuracy● To reconfirm the validity of fbBCI modality

1: Introduction - Research Purpose

9

● Test several signal preprocessing combinations ①○ Downsampling○ Epoch averaging

● Classify with three different machine learning methods ②○ SWLDA○ Linear SVM○ Non-linear SVM (Gaussian kernel)

2: Method - Conditions

10

CommandBrainwave① ②

2: Method - Signal Acquisition

11

● Event related potential (ERP) ○ captures 800 ms after an onset of vibrotactile stimulus ○ next converted to a feature vector using EEG potential

ex.) fs = 512 [Hz] ERPinterval = 800 [ms] = 0.8 [sec] Vlength = ceil(512・0.8) = 410

Vlength

VCh○○

p[0]

p[Vlength - 1]

Vlength = ceil( fs・ERPinterval)where fs [Hz] , ERPinterval [sec]

Ch○○

2: Method - Signal Preprocessing(1)

12

● Downsampling (nd)○ ERPs were decimated by

4 (128 Hz), 16 (32 Hz) or kept intact (512 Hz)

○ To reduce vector length Vlength

nd = 4 (128 Hz) nd = 16 (32 Hz)

Ch○○ Ch○○

13

● Epoch averaging (ne)○ ERPs were averaged using 5, 10

ERPs or no averaging○ To cancel background noise

ne = 1 ne = 10

Ch○○ Ch○○

2: Method - Signal Preprocessing(2)

● Concatenating all feature vectors

2: Method - Feature Extraction

ex.) fs = 128 [Hz] (nd = 4) Vlength = ceil(128・0.8) = 103

14

Vlength

VCz …

Vlength

VPz …

Vlength

VCP6

… … ……Vex.) VlengthALL = Vlength・8 = 103・8 = 824

VlengthALL

Ch1 Ch2 Ch8

● Machine learning methods○ SWLDA○ Linear SVM

… K(u,v’) = u v’○ Non-linear SVM (Gaussian)

… K(u,v’) = exp(-γ||u-v|| ) γ = 1/VlengthALL , c = 1

2: Method - Classification (1)

15

T

2

● Training the classifier

2: Method - Classification (2)

16

VT1

VT2

VlengthALL VlengthALL

VN1

VN2

VTmax

VNmax

VTmax = 60 / ne VNmax = 300 / ne

Classifier (2cls)

Non-Target Target

● Training the classifier

2: Method - Classification (2)

17

VT1

VT2

VlengthALL VlengthALL

VN1

VN2

Classifier (2cls)

VTmax

VNmax

VTmax = 60 / ne VNmax = 60 / ne

Random chooseas many as Tmax

}

Non-Target Target

● Evaluation with the trained classifier○ Same nd and ne were applied

2: Method - Classification (3)

18

VT1

VlengthALL

VTmax = 10 / ne

Target? orNon-Target? Classifier (2cls)

Test data

● SWLDA classification accuracies○ BEST: 57.48 % (nd = 4, ne = 1)

3: Results - SWLDA

19

Signal decimation (nd)

● Linear SVM classification accuracies○ BEST: 58.5 % (nd = 16, ne = 10)

3: Results - Linear SVM

20

Signal decimation (nd)

● Non-linear SVM classification accuracies○ BEST: 59.83 % (nd = 4, ne = 1)

3: Results - Non-linear SVM

21

Signal decimation (nd)

4: Discussion and conclusions

22

● fbBCI classification accuracy has been improved○ Both nd and ne combinations were tested○ 53.67 % in previous reported results

⇒ 59.83 % by non-linear SVM (nd = 4, ne = 1)○ 58.5 % by linear SVM and 57.48 % by SWLDA

● The potential validity of fbBCI modality was reconfirmed○ Expect to improve a QoL for ALS patients

● However, more analyses would be required○ Only 10 healthy users of fbBCI paradigm○ Need higher accuracies in case of a practical application

23

Many thanks for your attention!