![](bg7.png)
6.2.1 Principal Components Analysis
The data resulting from RFpat subsystem and Gardien
subsystem were analyzed simultaneously through
their respective variables: posture, activity, cardiac
frequency, exposure time under the infra-red sensor
C3 (sensor indicating the input/output of the person in
the home), and exposure time under the infra-red sen-
sor C5 (sensor indicating the fall of the person). The
analysis of the PCA algorithm results made it possible
to propose a set of decision rules on several levels:
• To define an estimator in two levels: a ”physio-
logical” distance between two parameters (cardiac
frequency, activity) normally correlated: normal
state if they are close or pathological state if they
are distant. Then, a ”actimetric” distance (Slope,
C5): normal if distant or pathological if close.
• There are a correlation between the cardiac fre-
quency and the activity which will allow the fu-
sion system to avoid a malfunction of one of the
two sensors.
6.2.2 Application of the Neural Networks
The Neural Networks (NN) consist in an input layer,
the sensors signals, several transition layers (denoted
as hidden layers) and of an output layer delivering
the classification of the data observed in situation ei-
ther ‘Normal’, or critical ‘detected Fall’. A classi-
cal NN structure was implemented by using a Multi-
Layer Perceptron (MLP) based on only one hidden
layer consisting in eight neurons after an optimal tun-
ing.
Each neuron realizes a scalar product between its
input vector and the weight vector, where a deviation
is added, then operates an activation function in order
to generate its output value y: y = f (x.w + b).
The activation function must be strictly crescent and
bounded. A classical function used in our experiment
is the standard sigmoid function whose equation is re-
minded hereafter: f (x) =
tanh(x)+1
2
.
Two types of networks were compared, with respec-
tively as input vector of the MLP first layer:
• Single actimetric data of RFpat in entry of the
network, giving a rate of recognition of the order
84%.
• The actimetric data of RFpat and horizontal infa-
red sensors of Gardien, providing a rate of 86%.
The improvement nevertheless remains quite limited.
One improvement track is to increase the data cor-
pus used for the learning phase, namely by recording
more specific actual and simulated emergency situa-
tions thanks to the multi-modal recording tool previ-
ously described in this paper. Another main improve-
ment track will be investigated by adding the AnaSon
(abnormal sound detector) modality. Therefore that
is why the need of a new multi-modal recording tool
was considered as crucial for the follow-up. Thor-
ough investigations will also be performed again on
KNN and GMM techniques, namely by working on
the data pre-processing (normalized, transformed in-
put data).
7 CONCLUSIONS
This paper has focused on the technology used for
implementing the acquisition step of the platform.
Preliminary results are encouraging with the achieve-
ment a multimodal medical database including pa-
tient’s clinical data, usual environment sounds and pa-
tient localization. The platform enables us to have a
full and tightly controlled universe of data sets and to
evaluate the decision part of remote monitoring sys-
tems.
Our platform is in the research phase targeting a pro-
totype, the system will be completed and improved by
adding a data fusion-based decision element exploit-
ing the measurements coming from this platform in
order to propose new processes to reinforce the secure
detection of patient’s distress events. In particular the
fall situations are studied: indeed one or more televig-
ilance modalities might be out of order, or a particular
environmental situation (ambiant noise, bad wireless
conditions, sensors disabilities .. .) can hide one par-
ticular modality or more. This is a very challenging
issue for hospital emergency units such as for instance
SAMU in France or Telecare services providers in
general. Studies of usability are planned, in order to
test the satisfactoriness of patients towards this system
and to get a standardization prototype for our plat-
form. This constitutes indeed, a first concrete step
before a prototype deployment. In actual situation,
evaluation and connection to smart home system are
also planed to be performed in the framework of a
new European project.
REFERENCES
Boudy, J., JL.Baldinger, F.Delavault, Farin, D., R.Andreao,
S.Torres-Muller, A.Serra, F.Rocaries, C.Dietrich,
F.Steenkeste, A., Dr.M.Baer, Dr.Ozguler, and D.Gaiti
(June 2006). Telemedecine for elderly patients at
home. In the TelePat project.ICOST conference ulster
Belfast.
B.Th
´
elot (June 2003). Resultats de l’enquete permanente
sur les accidents de la vie courante. In Reseau
A MULTIMODAL PLATFORM FOR DATABASE RECORDING AND ELDERLY PEOPLE MONITORING
391