IEEE Bio medical engineering 2016 Title and Abstract
STTP_POSTER
1. Performance of an EEG-based BCI communication system
using a modified probabilistic language model
Dustin G. Roberts, William Speier, Nand Chandravadia, Nader Pouratian
UCLA Department of Neurosurgery
Dustin G. Roberts
UCLA David Geffen School of Medicine
Email: dgroberts@mednet.ucla.edu
Phone: (714) 604-6096
Contact
1. Farwell L, Donchin E. (1988) Talking off the top of your head: toward a mental prosthesis utilizing event-related brain potentials. Electroencephalgr Clin Neurophysiol. 70, pp. 510-523.
2. Huggins JE, Wren PA, Gruis KL. (2011) What would brain-computer interface users want? Opinions and priorities of potential users with amyotrophic lateral sclerosis. Amyotroph Lateral Scler. 12, pp. 318-324.
3. Speier W, Arnold C, Deshpande A, Knall J, Pouratian N. (2015) Incorporating advanced language models into the P300 speller using particle filtering. J Neural Eng. 12(4), pp. 046018.
4. Speier W, Arnold C, Lu J, Taira RK, Pouratian N. (2012) Natural language processing with dynamic classification improves P300 speller accuracy and bit rate. J Neural Eng. 9(1), pp. 016004.
5. Kaufmann T, Schulz SM, Grunzinger C, Kubler A. (2011) Flashing characters with famous faces improves ERP-based brain-computer interface performance. J Neural Eng. 8(5), pp. 045016.
6. Lu J, Speier W, Hu X, Pouratian N. (2013) The effects of stimulus timing features on P300 speller performance. Clinical Neurophysiology. 124(2):306-314. doi:10.1016/j.clinph.2012.08.002.
7. Schalk g, McFarland D, Hinterberger T, Birbaumer N, Wolpaw J. (2004) BCI2000: A General- Purpose Brain-Computer Interface (BCI) System. IEEE Trans Biomed Eng. 51, pp. 1034-1043.
8. Townsend G, LaPallo B, Boulay C et al. (2010) A novel P300-based brain–computer interface stimulus presentation paradigm: Moving beyond rows and columns. Clinical Neurophysiology. 121(7):1109-1120.
doi:10.1016/j.clinph.2010.01.030.
References
The P300 speller is an electroencephalogram (EEG)-based brain-
computer interface (BCI) that allows one to communicate language
by detecting event related potentials and converting them into
virtual commands in the form of typing. This mode of
communication is well-suited for patients with “locked-in”
syndrome, as commonly seen in Amyotrophic Lateral Sclerosis
(ALS).[1]
• Non-invasive electrodes create undesirable signal noise; this
necessitates averaging over multiple trials and decreases
character selection rate.[2]
• Previous classifiers have used brute force approaches that
sample entire state spaces and are computationally intensive.
Incorporating language information into the classifier can
improve system speed and accuracy. [4]
• The use of a particle filtering algorithm (Figure 1) creates
probabilistic automata using sequence importance resampling
with data from EEG responses to model the English language.[3]
Overview
Study #1
Online analysis in 10 healthy subjects yielded an average typing
speed of 12.49 characters/minute with 94.50% accuracy for an
average information transfer rate of 57.84 bits/minute. A paired t-
test demonstrated significant improvement in typing speed
(p=0.0006) and bit rate (p=0.04) compared to the PF classier
without word completions. Accuracy was lower on average, but
the difference was not statistically significant (p=0.19).
Study #2
Online analysis in three patients with ALS yielded an average
typing speed of 10.31 characters/minute with 98.00% accuracy for
an average information transfer rate of 51.16 bits/minute.
Compared to the group of 10 healthy subjects, no statistically
significant difference was observed for typing speed (p=0.21), bit
rate (p=0.14) or accuracy (p=0.21), suggesting that ALS patients
perform equally well.
Methods and Materials
Particle filtering effectively incorporates language information into
the classifier for the P300 speller communication system.[3]
Incorporating a word completion function into the classifier
appears to further improve performance. A preliminary study in
ALS patients suggests that this patient population is equally
capable of using the system. In summary:
• All 10 healthy subjects were able to type faster and had a
higher bit rate using the WC function, and 6 out of 10 achieved
better or equal accuracy.
• In three ALS patients, there was no significant difference in
online performance compared to healthy subjects for the non-
WC system. Offline analysis also demonstrated a similar trend.
Discussion
• Optimization of stimulus modality. Proposed experiment: row-
column vs. checkerboard[8] vs. combinatorial flashing paradigms
• Optimization of visual stimulus. Combining other visual
paradigms with the famous faces paradigm, such as a character
zooming or color change method can possibly yield superior
results.
• Optimization of flash duration and interstimulus interval for
the current system. Performance is heavily dependent upon
these values.[2] Improved classifiers may allow for shorter
stimuli as any loss in signal resolution could potentially be
overcome by the inclusion of language priors[3].
• While the initial results in ALS patients are promising, a larger
study needs to be conducted to determine the performance of
the P300 speller in this population.
Results
Figure 1. Simplified automaton representing three words: “a,” “the,” and “to.”[3]
1. Sequences of character
illuminations are presented
as ‘famous faces’[5] on a 6x6
matrix in BCI 2000.[7]
2. When the target character is
illuminated, an evoked
response is elicited in the
user’s EEG signal.
3. The classifier determines the
targeted character using a set
of prior probabilities and
evidence from evoked
responses.
4. Results are relayed back to
the monitor as feedback to
the user.
Figure 2. Schematic of the P300 speller.
Figure 4. Study #1 – Differences in selection rate (left), accuracy (middle) and information
transfer rate (right) among healthy subjects.
Figure 5. Study #2 – Differences in selection rate (left), accuracy (middle) and information
transfer rate (right) between ALS patients and healthy subjects.
Offline Analysis:
Online Analysis:
Figure 2. Offline performance in two ALS patients compared to healthy subjects.
Future Directions
Study #1 – Compare performance of the particle filtering (PF)
classifier in healthy subjects with and without word completions
(WC) using a row-column (RC) stimulus modality and ‘famous
faces’ (FF) visual stimuli.
Study #2 – Compare performance of ALS patients with healthy
subjects using the original PF classifier (i.e. without word
completions), RC stimulus modality, and FF visual stimuli.
FF visual stimuli: ‘famous
faces’ are superimposed over
characters in the matrix. This
method was previously shown
to evoke a greater response
over other traditional
methods of intensification.[5]
RC stimulus modality: a
pseudo-random flashing
paradigm is presented in
sequences of rows and
columns with an interstimulus
interval of 25ms and stimulus
duration of 100ms.[6]
PF classifier: the PF algorithm
models the English language
via probabilistic automata
using a language corpus in
corroboration with evidence
from EEG response signals.[3]
WC function: a fraction of
particles move through the
model to states associated
with strings of characters that
represent completed words.
States with the highest
probability are presented as
options to the user.