Biomedical Engineering Reference
In-Depth Information
3)
Finally, the arm is stationary if,
P
(
O|
λ
s
)
> P
(
O|
λ
m
)
(5.54)
and is moving if,
P
(
O|
λ
m
)
>
P
(
O|
λ
s
)
(5.55)
To explicitly compute
P
(
O
|λ), we use the practical and efficient forward algorithm [
40
] de-
scribed earlier in this section.
The classification decision in (
5.53
) and (
5.54
) is too simplistic because it does not optimize
for overall classification performance, and does not account for possible desirable performance met-
rics. For example, it may be very important for an eventual modeling scheme to err on the side of
predicting arm motion (i.e., moving class). Therefore, our previous classification decision to include
the following classification boundary is modified to :
P O
P O
(
λ
λ
=
)
(5.56)
m
y
(
)
s
where
y
now no longer has to be strictly equal to 1. Note that by varying the value of
y
, we can
essentially tune classification performance to fit our particular requirements for such a classifier.
Moreover, optimization of the classifier is now no longer a function of the individual HMM evalu-
ation probabilities, but rather a function of overall classification performance.
In addition, we have determined previously through first- and second-order statistical analy-
sis that some of the 104 recorded neural channels contribute only negligibly to arm movement
prediction [
41
]. Therefore, in the experiments we not only vary the four parameters listed above,
but also the subset of neural channels that are used as temporal features in the segmentation process.
Table
5.1
lists the seven different subsets of neural channels that were used in our experiments.
Tables
5.2
and
5.3
report experimental results for different combinations of the four param-
eters and subsets of neural channels. These tables present representative segmentation results for a
large number of experiments.
From Table
2
, we see that the VQ-HMM can reasonable classify the arm gross dynamics
(i.e., moving or not). Unfortunately, even after exhausting many parameter combinations the results
were only able to reach a plateau in performance at about 87%. We also note that because a subset
of neural channels at the input yielded the best performance, some of the measured neural channels
may offer little information about the arm movement. These two observations help to motivate the
next section.
