Portable Wireless Device for Automated Agitation Detection - E-Healthcare Systems and Wireless Communications - page 252

that z is a training point having the label (-1).
While Y 1 assumes that z has the label (1). It is now
necessary to define a quantity that estimates the
confidence in each of the assumptions (z belongs
to class 1 or class -1).
Note that: to every S(X, Y) corresponds a
unique couple (D, w). This can be justified by the
fact that D is unique and depends on the training
points X, and w is a single valued function of
S, X and Y. Since the VC dimension is closely
related to its upper bound, its upper bound, VC max
will be used.
Definition. Given S(X,Y) and an unseen vector
z, the measure of confidence on the decision d(z)
by S(X,Y) is given by:

confident then the point is sent to SVM-2 and its
decision will be the final decision. Note that this
architecture is recursive and could be extended to
SVM-n, and the unknown point will go deep in
the architecture until the decision made by SVM-j
is confident and then its decision will be the final
one. In agitation detection application, two SVMs
were enough to classify all the points.

EXPERIMENTAL RESULTS

The following section summarizes the experimen-
tal results of the device and the detection algorithm.

Device

( ) = ( ) ( ) − ( )

C z d z VC

max

X Y , VC

max

X Y ,

S z − 1 S z 1

The device was tested using the prototype for sen-
sor accuracy and repeatability. For this purpose
the following experimental setups were made.

The proposed multi-level SVM is a two-class
classifier based on cascading two SVMs. The first
SVM (SVM-1) is trained by “easily classifiable”
points while the second one (SVM-2) is trained
by not “easily classifiable” points. An “easily
classifiable” point is defined as follows:
Definition. Let X be the set of training points
and x

Temperature Accuracy
and Repeatability

To test the accuracy of the RTD sensor, a tem-
perature calibrator from Omega was used. The
calibrator gives the ability to fix a temperature with
high accuracy. The RTD sensor is then inserted
in the calibrator, and the result extracted from
our device is compared with the value fixed on
the calibrator, this is repeated 3 times. It is worth
noting that the temperature range needed for the
device is skin temperature range which is limited
in extreme cases between 20 o C and 40 o C. The

X is one specific training point. x is said
to be easily classifiable iif C(x) > 0.
Note that the confidence is computed by
SVM-1. The output of this architecture is defined
as follows:
Definition. Let z be an input of unknown
class, then:

∈

− 1 d z ( ) < 0 and C z ( ) > 0

1 1

or d z ( ) < 0 and C z ( ) < 0

Table 1. Experimental results for the temperature
sensor

2 1

Class =

1 d z ( ) > 0 and C z ( ) > 0

1 1

or d z ( ) > 0 and C z ( ) < 0

2 1

Calibrator
Temp ( o C)

Measured
Temp ( o C)

Accuracy
(%)

Repeatability
(%)

25 24.85 99.40 99.98

An unknown point z is classified by SVM-1
and a measure of confidence is then generated.
If SVM-1 is confident of its decision then this
decision will be the final one. If SVM-1 is not

30 29.88 99.60 99.97

35 34.84 99.55 99.99

40 39.87 99.675 99.99