Biomedical Engineering Reference
In-Depth Information
TaBlE 5.3:
Results compared against different systems on same data
SINGLE LINEAR
MODEL
BIMODEL
WITH VQ-HMM
BIMODEL WITH
ICHMM
Classification %
-
87%
93%
CC (move)
0.75 ± 0.20
0.79 ± 0.27
0.86 ± 0.11
CC (rest)
0.06 ± 0.22
0.06 ± 0.24
0.03 ± 0.26
Although the standard receiver operator characteristic (ROC) curve can determine the thresh-
old for an appropriate performance (i.e., compromise between number of false alarms and false detec-
tions), it fails to convey an explicit relationship between the classes and the threshold when optimizing
performance. We use the true positives and true negatives of the respective classes to provide an
explicit interclass relationship. For nomenclature simplicity, we label these two curves the likelihood
ratio operating characteristic (LROC) curves because they represent the quantities in the likelihood
ratio. The “optimal” operating point on the LROC occurs when the two curves of the classes intersect
because this intersection represents equal classification for the two classes.
In Figure
5.12
, the LROC curves show that the ICHMM is a significant improvement in clas-
sification over our previous VQ-HMM classifier. This is evident from the equilibrium point showing
that the movement and rest classifications occur around 93% as opposed to 87% in our previous work
mentioned in the VQ-HMM section. In Figure
5.12
, we also see that the threshold is similar in the
training and testing (ζ = 1.0044; ζ = 1.0048, respectively), which show that the threshold can be reli-
ably set from the training data.
FIgURE 5.13:
Trajectory results compared.
