Brain-computer interfaces:
classifying imaginary movements
and effects of tDCS
Iulia Comşa
MRes Computational Neuroscience and Cognitive Robotics
Supervisors: Dr Saber SamiDr Dietmar Heinke
Presentation structure
An overview of brain-computer interfaces
Experiment 1: effects of tDCS on the EEG
Implementing a brain-computer interface with robotic feedback
Experiment 2: imagined movements (pilot study)
Brain-computer interfaces (BCIs)
What is a BCI?
“A communication system that does not depend on the brain’s normal output pathways of peripheral nerves and muscles” (Wolpaw et al., 2000)
In this project: BCIs based on motor imagery
The structure of a BCI
Wolpaw et al. (2002)
Brain imaging techniques for BCIs
Electroencephalography (EEG) Records electric potentials from the scalp
Advantages: Very good temporal resolution Comfortable and cost-efficient
Already on the market for home entertainment
http://www.biosemi.com/
Brain imaging techniques for BCIs
Transcranial direct current stimulation (tDCS)
Direct current applied to the brain
Induces changes in cortical excitability Anodal: increases excitability Cathodal: decreases excitability
http://www.neuroconn.de
Brain imaging techniques for BCIs
Transcranial direct current stimulation (tDCS)
Influences TMS-induced motor evoked responses in real or imagined movements
(Lang et al. 2004, Quartarone et al. 2004)
Potential benefit for classification
No study in literature about its effect on the EEG in the motor area
http://www.neuroconn.de
Investigating the effects of tDCS
Question: Does tDCS produce significant changes in event-related potentials in the motor area?
Event-related potential (ERP): brief change in electric potential that follows a motor, sensory or cognitive event
Luck et al. (2007)
Investigating the effects of tDCS Previously collected data available
Three groups of participants (9 participants each) Anodal tDCS Cathodal tDCS Sham
Task 250 real finger taps 250 imaginary finger taps Two sessions: before and after tDCS
Data collection 128 EEG channels using a Biosemi ActiveTwo system
Investigating the effects of tDCS Data pre-processing (EEGLAB
Toolbox)
Filtering Between 1 and 100 Hz
Epochs (segments of data) were extracted between 0 and 1 second following the stimulus
Artefact rejection Removing data contaminated by noise (e.g.
blinks) By amplitude threshold (55-125 mV) and
manually
Investigating the effects of tDCS
Real taps
Anode
Cathode
Sham
Imagined taps
ERP grand averages (ERPLAB Toolbox)
Investigating the effects of tDCS Permutation t-tests (Mass Univariate ERP Toolbox)
Family-wise alpha level: 0.05
2500 permutations
Tmax statistic (Blair & Karniski, 1993)
Anode-Cathode t-scores, real finger taps after tDCS [video]
Investigating the effects of tDCS Significant differences for real taps
Anode-Cathode
Anode-Sham
Cathode-Sham
~ 85 ms
~ 230 ms
Differences for imagined taps
Investigating the effects of tDCS
Anode-Cathode
Anode-Sham
Cathode-Sham
~ 80 ms
~ 700 ms
Effects of tDCS on ERPs: Summary
Significant effects found for anodal tDCS in the motor area around 85 and 230 ms during real movements
Significant effects found for cathodal tDCS around 700 ms in the parietal area during imaginary movements
Although not always significant, differences in the motor area are visible in all conditions
Oscillatory EEG processes ERPs: phase-locked activity What if the response is not phase-locked?
Induced responses: EEG frequency bands Mu rhythms: 8-13 Hz
Recorded from the sensorimotor cortex while it is idle Briefly suppressed when an action is performed or
imagined
Beta rhythms: 13-30 Hz
Gamma rhythms: 30-40 Hz, 60-90 Hz
Building a BCI with robotic feedback
BCI2000a general-purpose system for BCI research consisting of
configurable modules
Signal Acquisition
StimulusPresentation
Signal Processing
BCILAB Toolbox - provides:•Signal preprocessing (filtering, cleaning)•Feature extraction: Common Spatial Patterns•Machine learning algorithms for classification
RWTH Aachen MINDSTORMS NXT Toolbox• Robot arm control
Imagined movements pilot study
3 healthy participants
Imagined left and right hand clenching
(100 trials each)
Data collection: 32 electrodes
covering the motor-premotor area
(using a Biosemi ActiveTwo system)
Imagined movements pilot study r2 (coefficient of determination): the amount of
variance that is accounted for by the task condition
Strongest activity: 10-30 Hz in lateral electrodes Some activity above 60 Hz
Participant 1 Participant 2 Participant 3
Channel
Frequency (1-70 Hz)
Imagined movements pilot study Best results – 10 fold cross-validation:
Epochs between 1 and 2 seconds after stimulus
Classifier: linear discriminant analysis
Participant 2: 88,5% accuracy Common Spatial Patterns FIR Filter: 10-30 Hz bandpass
Participant 3: 85,5% accuracy Filter-Bank Common Spatial Patterns Frequency windows: 8-30 Hz and 8-15 Hz
No model with accuracy better than 65% could be trained for Participant 1
Further work: Improving the results More trials
Problem: subjects may get bored
Adding online feedback Problem: we would already need a good classifier
Incorporating purpose in the motor imagery “Clenching a fist” versus “grabbing a pen”
Using tDCS 99% accuracy for the tDCS data from Experiment
1
Project summary
We showed that tDCS has significant effects on event-related potentials
We implemented a brain-computer interface with robotic feedback
We performed a pilot study and explored classification of left and right imaginary movements
Thank you.