Portable Wireless Device for Automated Agitation Detection - E-Healthcare Systems and Wireless Communications - page 254

(show a block of color only), second set is called
congruent word slides (word and color matches),
third set is called incongruent word slides (words
do not match color presented). The test included
60 randomly set color blocks (1min) where sub-
jects are to name out loud the color they see, 60
randomly set congruent word slides (1min) where
subjects are to read out loud the word they see, and
finally 120 randomly set incongruent word slides
(2min) where subjects are to name the color of
the word they see (not read the word). Each slide
is set to show for one second. If subjects missed
a slide they were asked to move on to the next
one. The overall objective of the subject was to
get as many correct answers as they can. Subjects
were asked to complete the Trait Scale State-Trait
Anxiety Inventory (T-STAI) before and after the
test. The two STAI filled up by a subject allowed
us to validate their state.
Data from subjects that demonstrated stress
using the T-STAI analysis was used to evaluate
the accuracy of the detection algorithms. Each
patient had 240 samples (RTD, GSR, IBI) taken
synchronously corresponding to the 3 Stroop
test slides. In total 58 subjects were tested. It is
worth noting that in order to test the architecture
efficiently in a statistically significant way the
following results correspond to computations
done off line. The K-fold cross-validation tech-
nique was used. The 58 subjects are subdivided
into groups of 2 which yields 29 folds. In order
to show the robustness of this architecture against
over fitting, only 1 fold is taken for training and
the remaining folds are used for validation. From
the training fold, a subset X of 80 triplets (ST,
GSR, IBI) is chosen randomly. X is used to train a
traditional single SVM using the Gaussian kernel
function, and all the samples of the remaining
28 folds were used to test its accuracy. The same
set X was split into two sets X 1 and X 2 . SVM-1
was trained using X 1 while SVM-2 was trained
using X 2 and the samples from the remaining
28 folds were used in the testing procedure. The

comparison of the performance of a single SVM
with the performance of the multi-level SVM is
valid because in both cases, the same training set
X is used. The average accuracy for the proposed
architecture reached 91.4% with a standard de-
viation of 1.71% while the average accuracy for
the normal SVM reached 90.9% with a standard
deviation of 1.94%.

CHALLENGES AND
FUTURE RESEARCH

The challenges for portable and intelligent devices
can be divided into technical and non-technical.
As any portable technology, size and weight are
a primary concern. For medical applications, this
constraint is more stringent as these would be
devices used by the elderly or weaker subjects.
In addition, the operational life of the device
and its power consumption are a concern. Such
devices would be expected to operate for years
with minimal intervention. Human interfacing is
another major challenge as the device must be
“transparent” to the subject with no impact on
daily activities. This is where the role of wire-
less technology becomes even more important in
reducing the wires running along the body of the
subject. Obviously, all this has to be accomplished
without compromising the performance of the
device in the measurements and decisions taken. A
challenge specific to devices detecting a subject's
state is ground truth definition. For example, in
the case of agitation detection, how can we define
the exact point of agitation onset even when the
subject is under direct observation? The standard
in human subject testing is to define this point at
the introduction of a stressor. However, this is not
always the exact point as some subjects might be
exhibiting agitation symptoms prior to the stressor
introduction or vice versa. On a different note,
clearly whenever a system is introduced to the
medical field, especially one that autonomously