Computational Intelligence - February 2016 - 34

20
0

20
40
CDT (s)

0.9
0.8
0.7
0.6
0.5

Short

Long

sb2, MDT: 3.2 s

80
60
40
20
0

5
10
CDT (s)

Average Classifier Output

sb1

1.0

Percentage of Trials

Average Classifier Output

Percentage of Trials

sb1, MDT: 2.7 s

sb2

1.0
0.9
0.8
0.7
0.6
0.5

Short

Long

Figure 1 Distribution of command delivery time (CDT) for two subjects (sb1 and sb2) in an MI online protocol. The comparison of short and
long trials (with respect to the median delivery time, MDT) reveals significant differences in the average classifier output ^ p 1 0.001h .

Having long CDT is usually frustrating for subjects and may
reduce their engagement in performing the mental task. In addition, it can increase the workload or affect the performance of the
system if the BCI application has to meet temporal constraints. For
example, consider the task of controlling a brain-actuated robot
[19]. In this case, the user delivers right or left commands, while
the robot moves forward if the subject voluntarily decides not to
deliver any command. To make the robot cross a doorway on the
right side, the user needs to be fast enough to deliver the corresponding command at the proper moment. Otherwise, the robot
will miss the doorway and the user will need to deliver additional
commands to bring it back. In such cases, predicting that a command is going to take a longer time to be delivered would be
extremely helpful, as it would enable providing adequate assistance
to the user. For instance, the speed of the robot can be reduced so
that the user has enough time to make it cross the doorway.
As a result, adding a performance estimator to the BCI system
may allow to compensate for performance variations by adapting
the interaction to the user's current capabilities (e.g., through the
use of shared control) [21]. Figure 2 shows such a system that works
in parallel with the BCI and modulates action generation based on
an estimation of the performance. The goal of this study is to propose a method capable of making a trial-by-trial prediction of the
performance in an MI-BCI in terms of the time it takes to deliver a
command. This predictor estimates the reliability of a command
(i.e., having 'short' or 'long' CDT) based on the input features to the
BCI. For this approach to be useful,
the estimation has to be performed
at the very beginning of a mental
Performance
task execution. Here, we show that
Estimation
the information within 1.5 s from
the beginning of the MI execution
is sufficient to make such a predicFeature
Classification
Action
Extraction
tion. Importantly, the experiments
Generation
were done in an online setting lastFeedback
ing several sessions (2-3 depending
on the subject) so as to account for
the effect of feedback on the modFigure 2 Common block diagram of a BCI in the dashed box. The goal is to design a performance
ulation of brain signals. Finally, the
estimator, which works in parallel with the BCI and can be used to provide online adaptive assistance
feasibility of using this method for
for the user. In this study, performance estimation is achieved by a trial-by-trial prediction of command
online adaptation of assistance has
delivery time (CDT).

follows that it is essential to evaluate performance changes in
online sessions, as distribution of data normally differs between
offline and online sessions because of the feedback subjects
receive in the latter. Second, one of the challenges for a BCI is
to cope with performance variations over extended periods of
time (e.g., over different sessions/days). Hence, a method capable of evaluating these variations over different sessions would
improve usability and reliability. Third, it is crucial to predict
performance on a short-time basis so as to compensate for the
user's varying capabilities while using a brain-controlled device.
BCIs based on motor imagery (MI-BCIs) typically combine
the outcome of classification for samples in a given time window in order to improve reliability. Some implementations
make a decision after a fixed amount of time [2]. Others accumulate evidence over time by integrating the classifier outputs
until it reaches a certain threshold, at which point the command
is delivered [18]. The latter approach results in different command delivery time (CDT) across trials [19, 20]. Figure 1 illustrates such variations in a BCI session for two subjects. Since the
integrated probability should reach a threshold for a command
to be delivered, long CDT should be due to lower average classifier output for samples of the trial. Indeed, after separating the
trials into short and long ones based on the median delivery
time (MDT), we observe that the average classifier output across
samples is higher for short trials compared to long ones
(p < 0.001 in Wilcoxon rank sum test, Figure 1).

IEEE ComputatIonal IntEllIgEnCE magazInE | FEbruary 2016

Table of Contents for the Digital Edition of Computational Intelligence - February 2016