Movement-Related Potential
- MRPs are small and slow potential drift before a voluntary movement
- the potentials are different between left and right arm, so it's easy to distinguish left vs right movement intention.
- harder detection than alpha and beta rhythms
- however, can be more specific to supplementary and primary motor cortical areas
Experimentation
- Hiraiwa (1990)- backpropagation neural network to test voluntary utterance of vowels (a, e, i, o, u) and moving the joystick in 4 directions
- 53% patterns correctly classified for vowels, 96% patterns recognized by directions
- vowels require a higher level of cognitive activity, hence a harder pattern recognition
- asyncrhonous switch- BCI can detect from a user is idle to active control state
- low-frequency AS- record 1-4 Hz range
More Experimentation!
- Shenoy and Rao (2005) - dynamic Bayesian network
- probability distributions over brain and body-states during planning + execution
- continuous tracking, allows generation of control signals
- MRPs can estimate movement
Stimulus-Evoked Potentials
- stereotypical EEG responses but by specific stimuli
- rare stimulus triggers a P300 response (300 ms)
- The P300 response stimulus itself is unpredictable, but is always 300 ms after the stimulus.
- Generally in the parietal lobe, but sometimes certain components in the temporal and frontal lobes as well.
- oddball paradigm- spelling a word using a matrix of letters and are flashed in random orders.
Every time a letter flashed, the brain will generate a P300 reponse.
Steady State Visually Evoked Potential
- detecting SSVEP caused by a continuously fluctuating stimuli (hence the name steady state)
- Ex: decoding one of two possible choices, these two choices can be represented by a light at different frequency. Then we can see the difference in stimuli.
- Each button was lit up at a different frequency ranging from 6-14 Hz. Therefore, decoding SSVEP is possible.
- 8 out of 13 subjects used SSVEP subject to type desire phone number. Avg: 27.15 bpm for ITR.
Auditory Evoked Potentials
- Donchin and colleagues- used P300 in spoken commands (yes, no, pass, and end), achieved accuracy in 63%-80% in 3 ALS pateitns
- Hill and colleagues (2005)- ICA with support vector machines. 50 ms square-wave beeps of different frequencies either on the left or right ear. Subjects count the amount of beep on the left or right side.
More on Auditory Evoked Potentials
- Furdea and colleagues: letters in matrix coded with acoustically presented numbers
- Less accurate than a visual BCI- Halder (2010) found that ITR has an avg of 2.46 bpm and accuracy around 78.5%.
- Only reliable to patients who have lost all motor function and have short attention spans.
BCIs based on Cognitive Tasks
- requires higher levels of thinking, but each cognitive task can be mapped to a control signal
- can discriminate between activity patterns of different tasks
- Anderson (1996)- (1) relax, (2) letter composition (3) math, (4) visual counting, (5) geometric figure
- EEG recorded from 10-20 system. Autoregressive model with two/three-layer backpropagation neural networks
BCIs based on Cognitive Tasks
- average accuracy ranged from 71% to 38%
- Galan (2008) - 3 mental tasks to operate simulated wheelchair along specified path. (1) mentally search for words, (2) relaxing while fixating on the center of screen, (3) motor imagery of left hand
- used LDA classifier. 100% and 80% (different subjects) of the final goals.
Error Potentials in BCIs
- error detection is very important
- detection is done by recognizing the brain's reaction to error: hence the error potential
- Buttfield (2006) - deliberately make 20% errors to record ErrPs from frontocentral region.
- with errors - accuracy 79.9%, without errors- accuracy 82.4%
Coadaptive BCI
- BCI can adapt to a user's brain signals continuously
- non-stationary learning task - continuously mapping brain signals to control signals
- 3-6 minutes of adaptation go a long way
- Buttfield (2006) - continuously adapting mean and covariance to calculate the learning rates and online adaptation is better than static (for up to 20.3%)
- DiGiovanna (2009) tried operant condition with BCI
Hierarchy BCI
- Continuous non-invasive BCI might be cognitively taxing
- hierarchical BCI - user teaches BCI system new skills, then invoked directly as high level command
- developed to control a humanoid robot
- hierarchical control is faster and more accurate
Other Non-Invasive BCIs
fMRI
- Can a subject learn to control their blood oxygen level dependent (BOLD) response?
- Weiskopf (2003) - feedback paradigm- local BOLD signals constantly provided with a delay < 2 sec
- subjects can increase or decrease BOLD responses
Magnetoencephalography (MEG)
- higher spatiotemporal resolution than EEG
- Mellinger (2007) - investigated a MEG on voluntary amplitude modulation of sensorimotor mu and beta rhythm
- spatial filtering method based on geometric properties of signal propagation
- successfully communicate limb movements
fNIR
- captures hemodynamic response
- Coyle (2004) - fNIR detects characteristic repsonses when subjects engage in motor imagery and utilize this response to control an application
- easy for drawing and cursor control
- however, average completion time actually decreased with practice
Movement-Related Potential MRPs are small and slow potential drift before a voluntary movement the potentials are different between left and right arm, so it's easy to distinguish left vs right movement intention. harder detection than alpha and beta rhythms however, can be more specific to supplementary and primary motor cortical areas