Big Data Approach for Managing the Information from Genomics, Proteomics, and Wireless Sensing in E-health

Abstract

This chapter aims to show that big data techniques can serve for dealing with the information coming from medical signal devices such as bio-arrays, electro-physiologic recorders, mass spectrometers and wireless sensors in e-health applications, in which data fusion is needed for the personalization of Internet services allowing chronic patients, such as patients suffering cardio-respiratory diseases, to be monitored and educated in order to maintain a comfortable lifestyle at home or at their place of life. Therefore, after describing the main tools available in the big data approach for analyzing and interpreting data, several examples of medical signal devices are presented, such as physiologic recorders and actimetric sensors used to monitor a person at home. The information provided by the pathologic profiles detected and clustered thanks to big data algorithms, is exploited to calibrate the surveillance at home, personalize alarms and give adapted preventive and therapeutic education.

Full text

Chapter 1
Big data approach for managing the information
from genomics, proteomics, and wireless sens-
ing in e-Health
J. Demongeot, M. Jelassi, C.Taramasco
AGEIS, EA 7407, Faculty of Medicine, University Grenoble Alpes, 38700 La
Tronche, France, Jacques.Demongeot@agim.eu
RIADI, National Engineering School of Computer Sciences, Manouba Univer-
sity, 2010 Manouba, Tunisia, jlassi.mariem.esti@gmail.com
Escuela de Ingeniería Civil en Informática, Universidad de Valparaíso, General
Cruz 222, Valparaíso, Chile, Carla.Taramasco@uv.cl
Abstract. This chapter aims to show that big data techniques can
serve for dealing with the information coming from medical signal
devices such as bio-arrays, electro-physiologic recorders, mass spec-
trometers and wireless sensors in e-health applications, in which da-
ta fusion is needed for the personalization of Internet services allow-
ing chronic patients, such as patients suffering cardio-respiratory
diseases, to be monitored and educated in order to maintain a com-
fortable lifestyle at home or at his place of life. Therefore, after de-
scribing the main tools available in the big data approach for analyz-
ing and interpreting data, several examples of medical signal devices
are presented, such as physiologic recorders and actimetric sensors
used to monitor a person at home. The information provided by the
pathologic profiles detected and clustered thanks to big data algo-
rithms, is exploited to calibrate the surveillance at home, personalize
alarms and give adapted preventive and therapeutic education.
Keywords. Big data, genomics, proteomics, wireless sensing, e-
health, data fusion, alarm triggering
2
1
2
1
3
3

2 J. Demongeot, M. Jelassi, C.Taramasco
1.1 Introduction
The recent big data techniques are useful for compressing, clustering and
modeling health data during the visualization, analysis and interpretation
processes, especially in medical practice. The flow of data coming from e-
health wireless sensing, namely from the systems of surveillance of elderly
or chronic patients at home is dramatically increasing and needs to be
treated thanks to the “new” big data tools (often a new combination of
former classical approaches) (Figure 1).
Figure 1 Medical data produced pro year in a hospital receiving a total
amount of 200 000 patients / year.
The present overflow of health data has as consequence to displace the
barycenter of the information from hospital centered systems (such as the
first developed in Europe, Diogène® in Geneva, Bazis® in Leyden and
Crystal Net® in Grenoble) to the patient. This tendency is also observed in
USA, where information systems like the e-blue button by Humetrix® are
patient-centered and at least so used by paramedics and by the patients
themselves at home, than by physicians. The information content of the
medical file of a given patient was in the seventies about 500 ko of clinical
and biological data. With the development of computerized medical imag-
ing devices and their storage facilities (such as PACS, Picture Archiving
and Communication Systems) in the eighties, this information content
passed to 5Mo, but, with the introduction of modern tools of genomic and
Proteins
SNPs
0.1
2

3
proteomic studies (namely bio-arrays and mass or NMR spectroscopy),
this content jumped to 5 Go and more. The genome indeed is coding for
about 20,000 structural and metabolic proteins, plus about 230,000 immu-
no-proteins. By considering the regulatory networks in which genes and
proteins are involved, even depending in a binary way on their concentra-
tion (reduced to the value 1 over a certain threshold of expression and 0
under this threshold), it is possible to manipulate trajectories of data along
the duration of a disease in a huge state space, the Boolean hypercube
{0,1}250,000, whose size is 1075,000. These trajectories are depending on a set
of dynamic rules relating the state of a patient at time t to his state at time
t+1. These dynamic rules are closely depending on the existence of rela-
tionships of various types (inhibitory or activatory) between the genes and
proteins of the regulatory metabolic networks directly linked to the healthy
or pathologic state of the patient.
Section 1.2. consists of a rapid survey of the big data approach in
health, from the start of the usage of numerical data in medicine to the in-
vention of the main big data operators, such as the Neural Network (NN)
convolution and renormalization, used now for treating new e-health data.
The history of these tools is simple: after a multiplication of data analysis
tools from the seventies until the nineties (such as Spearman’s factorial
analysis, Benzécri’s correspondence analysis, MacQueen’s k-means, Di-
day’s “nuées dynamiques”, Hopfield’s neural networks, Kohonen’s maps,
Hérault/Jutten/Comon’s Independent Component Analysis, etc.), the suc-
cessive introduction of powerful storage tools (from hierarchical to rela-
tional data base management systems, and from object-oriented to no-SQL
data warehouses) pushed the medical data provider and user to invent new
analysis tools for managing and processing the stored big data, these tools
being often just a smart combination of previous mathematical algorithms
of data descriptive analysis (mixed if necessary to classical tools of infer-
ential statistics). After this survey, examples of recently invented big data
algorithms are described, presenting successively a big data compression
tool generalizing the wavelet transform (called “dynalet” transform), then
a weak version of the order statistics (called “monotony signature”), after a
clustering method, the Markovian final classes method (known as MCL),
useful for clustering cardio-respiratory data coming from the surveillance
at home of patients suffering cardio-respiratory diseases, and eventually, a
method of modeling using an improvement of the classical kernel estima-
tion method of the functional non-parametric approach.

4 J. Demongeot, M. Jelassi, C.Taramasco
The Section 1.4. describes some redundant multi-sensing techniques
necessary for example to anticipate changes in the ordinary life of a person
at home in terms of i) actimetry for detecting for example the appearance
of a progressive motor impairment, ii) sensory control for measuring the
loss of visual or auditory abilities and iii) vegetative surveillance for fol-
lowing the main physiological cardio-respiratory variables. The third Sec-
tion is dedicated to an example where critical profiles and alarms are built
using the big data methodology. This example concerns the detection in a
population at risk of pulmonary edema or cardiac failure, in which the part
of the vegetative nervous system made of the union at the bulbar level of
the respiratory and cardio-moderator centers is represented by a couple of
two 2-order differential systems of Liénard type, whose parameter values
at rest have been used for building an a priori classification of the cardio-
respiratory profiles.
In the Conclusion, the present big data approach are compared with the
complex systems one with which it shares many concepts, especially those
related to emergent properties due to the combination of a mass of infor-
mation coming from various redundant sources related to different interac-
tion networks, like the social, genetic, metabolic and physiologic ones, all
necessary to implement effectively the predictive medicine of the future.
1.2 The Big data approach
1.2.1 Ancestors of the digitization of health data
The first usage of numeric tables in medicine has been proposed by Abū
Bakr Muhammad ibn Zakariyyā al-Rāzī (865-935), a Persian polymath and
physician called also Muhammad Rhazes [1], who probably wrote the
book “The Secret of Secrets" [2, 3], a pseudo-Aristotelian treatise in form
of a letter from Aristotle to his student Alexander the Great treating a wide
range of topics, including astrology and medicine, and containing two
charts from an Arabic copy for determining whether a patient will recover
and survive, or remain frail and die based partly on the numerical value of
the letters of his name (Figure 2). The old dream of the “Secret of Secrets”
will be realized in the future by the predictive medicine, whose aim is to
predict for a given patient the destiny of his diseases, even before their
phenotypic occurrence.
The founder of algebra in 840, al-Khwârizmî and his numerous disciples
like Guillaume d’Auberive (1120-1180) and his Cistercian colleagues
Odon de Morimond and Thibaut de Langres [4], or Shaiykh Bahā’ī (1546-

5
1622) [5] knew the science of numbers and the charts of the Secret of Se-
crets, which is an amazing mix of simple rules, such as arithmetic se-
quences and some sophisticated rules as the triangle of W.T. Tutte [6]
(Figure 2), whose properties were surely ignored by Rhazes. It was only in
2010 that the smallest perfect equilateral triangle dissection by equilateral
triangles has been exhibited [7], as conjectured by W.T. Tutte. Numbers of
the right chart of Figure 2 are inside white external (11, 19, 20) and inter-
nal (2, 3) sub-triangles of the Tutte’s triangle, or are sums (modulo 30) of
the numbers inside these sub-triangles related by red links. Remaining
numbers between 1 and 30 belong to the left chart and represent arithmetic
series, the first starting at 6 with step 3 (except 12, badly transcribed in 28
on the right chart), the second starting at 5 step 5 (except 20, replaced by
22, a cabalistic number). All these rules constitute a necessary and suffi-
cient condition for belonging to the charts.
The epidemiological approach of the contagious diseases spread using
real data is more recent. The first example concerns both social and infec-
tious contagious diseases: St Anthony Monastery near Grenoble was in
charge of watching and curing infectious spreads like the Black Death of
1348 (Figure 3) and social epidemics like the St Anthony’s fire disease
(called now ergotism, also known as ergot poisoning). The Antonin con-
gregation was founded in 1095 by Gaston de Valloire, a nobleman of the
Dauphiné, and confirmed by the Pope Urban II the same year in thanksgiv-
ing for his son's miraculous cure from St. Anthony's fire thanks to the rel-
ics of Saint Anthony the Great. During the XIVth and XVth centuries, the St
Anthony’s monastic Order had 640 hospitals and 10,000 brothers in Eu-
rope along the roads to Santiago de Compostela and Jerusalem watching
plague, leper and ergotism. This monastery is also known for having shel-
tered two illustrious mathematicians, Jean Borrel (1492-1564), called also
Johannes Buteo, an algebraist who refuted the circle quadrature theory [8]
and Jean Duchon, inventor of the thin plates splines theory [9-11].

6 J. Demongeot, M. Jelassi, C.Taramasco
Figure 2 Top: two charts from “The Secret of Secrets” [2, 3], the left (re-
spectively right) chart corresponding to frailty and death (respectively
healing and life). Bottom left: arithmetic series by Guillaume d’Auberive
(bottom) [4] and Shaiykh Bahā’ī (top) [5]. Bottom right: smallest perfect
equilateral Tutte’s triangle dissection by small equilateral sub-triangles [6].
The Black Death (plague) watched by St Anthony’s Order has been the
major epidemic of the Middle Age and its demographic and socio-
economic consequences were dramatic. A mean estimation counts for the
half the part of the occidental population killed during the epidemic wave
(that is between 75 and 100 million of persons…). Born in the Caspian sea
area, the epidemic wave went through the Mediterranean commercial
routes. It reached ports like Marseilles in France and Genoa in Italy at the
end of 1347, during 5 years was spread widely in Europe and come back to

7
the Caspian reservoir. A simple Susceptible-Infected-Recovered (SIR)
model with viscosity proportional to altitude explains the essential of the
front wave dynamics observed during the half decade 1347-1352 (Figure
3). The model uses only 3 coefficients, a local diffusion coefficient in-
versely proportional to the altitude, a contagion parameter and a recovering
rate (representing the immunization rate after cure of the plague disease),
after the approach proposed at Geneva by Daniel Bernoulli in 1760 [12]
and at Paris by Jean Le Rond d’Alembert (the son of the famous Madame
du Tencin, close friend of Voltaire living near Grenoble) in 1761 [13].
They first discovered the quadratic equation of epidemic spread. Despite
its simplicity, the Black Death model is able to render count qualitatively
for the morphology of the front waves and, after calibrating the time on the
first front, for their velocity. Despite the quarantine measure decided after
the Black Death epidemics, the plague arrived anew in 1720 at the port of
Marseilles on a merchant ship, called the Grand-Saint-Antoine… More
than the surveillance of infectious (plague) or social (ergotism) contagious
diseases, the St. Anthony Order inaugurated the care of the sick at home.
Five centuries after, St Camillus de Lellis (1550 -1614), an Italian priest
founded an Order, dedicated to the care of sicks at home. Following Anto-
nin tradition, he created the first dedicated hospices (1584), where he
mixed people of all ages and conditions, as precursor of the intergenera-
tional homes. In 1777, this tradition was lost in Dauphiné, when St. An-
thony Order was transferred to the Order of Saint John of Jerusalem (cur-
rently Order of Malta), but French encyclopedists restored the medieval
vision of man as a whole, especially the “Physiocrates”, ancestors of the
physiologists and early supporters of the health monitoring and education
at home, allowed now in the 5P Medicine (Personalized, Preventive, Par-
ticipative, Predictive and Pluri-expert) by the tools recording and storing
health data, from Internet of Things (IoT), hospital data warehouses and
big genomic and proteomic repositories.

8 J. Demongeot, M. Jelassi, C.Taramasco
Figure 3 Left: simulations showing the front wave of the epidemic spread
(from Marseille and Genoa) after 3 months (middle) and 6 months (bot-
tom) with a diffusion coefficient inversely proportional to the altitude.
Right: Black Death spread showing the real front waves of 1348 (red),
1349 (blue) and 1350 (green), as observed by St Anthony’s monastery
(Left top thumbnail).
1.2.2 The main big data operators: Neural Network convolu-
tion, renormalization, gradient back-propagation and
simulated annealing
Neural Network (NN) approach is the core of the modern version of the
machine learning called the “deep learning”. After a first classical phase
where neural networks were used for treating raw data, mimicking some
biological operators like the lateral inhibition in retina on the visual way
[14, 15] or in cochlear nucleus on the auditory way [16], they were de-
clined on various form for learning tasks [17, 18]. More recently, the in-
vention of the “deep learning” approach has been proposed, often by peo-
ple like G. Hinton [19], Y. LeCun [18, 19] and P. Gallinari [20] already
deeply involved in the first classical phase. The main operators on neural
networks used in deep learning, i.e., discrete convolution, network renor-
malization, gradient back-propagation and simulated annealing, have been
already proposed either in a continuous mathematical formulation (1952
by L. Schwartz, for the convolution of distributions: “pour régulariser, on
convole”), or in a physical framework (1972 by K.G. Wilson, for the group

9
of renormalization), in image processing (1984 by D. and S. Geman, for
the simulated annealing) and in recognition tasks (1986 by D. Rumelhart,
G. Hinton and R. Williams, for the gradient back-propagation). The main
novelty in deep learning is the successive use of these operators for finding
hidden relationships between the entities providing data “tsunamis”, e.g.,
caused in medical applications by the multiplication of new medical in-
formation sources (from computerized imaging, protein spectroscopy and
genome identification tools, to the wireless multi-sensing sources of e-
health).
An example of deep learning architecture is the LeNet-5 network by Y.
LeCun [19]. The Figure 4 shows how the sequential use of various opera-
tors can extract hidden features from a mass of data, related for example to
a genetic network. The initial observation concerns DNA-array data, from
which it is decided if a gene is expressing a protein (state 1, corresponding
to the protein concentration over a certain threshold) or not (state 0, corre-
sponding to the protein concentration under this threshold). The succession
of these binary states in time, e.g. 120 successive records (four blood sam-
ples each day during a month) in a big study of 10,000 patients suffering
the same chronic disease, thanks to a very high density DNA microarray
(e.g., with 10,000 single-nucleotide polymorphism SNP’s as probes) gives
a part of the trajectory of gene expression configurations with 1010 data.
The aim of the interpretation of such data is double: i) identify the final
configurations (called attractors) of the dynamics of expression and ii)
confirm the architecture of the genetic network by identifying inhibitions
and activations (respectively red and black arrows on Figure 4 Top left). If
statistical procedures (mainly based on correlation analysis, in a spatio-
temporal Markovian or renewal context [14-17], but with the risk of get-
ting parasitic interactions, due to purely statistical and non causal correla-
tions) are giving too partial information about the interactions between
genes and their nature (inhibitory or activatory), the deep learning ap-
proach can be used for clustering observed configurations inside classes of
states corresponding to the attraction basins of the dynamics of expression,
the attraction basin of an attractor being the set of configuration trajecto-
ries having final configurations belonging to this attractor (Figure 4 Bot-
tom middle). Of course, if the number of genes is reasonable (about 100,
with for example the half located in strong connected components of the
interaction graph of the genetic regulatory network) and if all interactions
are known in the genetic networks, then a direct procedure of simulation of
the Boolean dynamical system underlying the dynamics, allows for obtain-
ing their attractors and attraction basins. If this direct approach (left blue

10 J. Demongeot, M. Jelassi, C.Taramasco
arrow on Figure 4 Bottom left) is impossible (e.g., due to the ignorance on
the nature of the interactions), the clustering operator at the end of the deep
learning procedure can solve the problem ii) of finding attractors and ba-
sins (blue arrow on Figure 4 Bottom right) and after, by solving the inverse
problem (red arrow on Figure 4 Bottom left), give indications for complet-
ing the interaction graph of the genetic network.
Figure 4 Top left: genetic network whose vertices are genes in state 1 (ex-
pressing proteins) or 0 (not expressing). Top middle: stack of several neu-
ral networks encoding each a part of the information coming from the ob-
servation of the genetic network states, processed by intra-network
operators, like lateral inhibition for contrasting and renormalization for av-
eraging. Top right: subsampling and clustering using inter-network opera-
tors like convolution. Bottom: clustering in classes corresponding to the fi-
nal configurations of the expression trajectories (expression attractors).
encoding

11
1.2.3. Big data approaches in health
Many useful big data techniques have been used recently for treating
health data, coming from the IoT [21, 22] or from image mining in big
medical image repositories or in hospital PACS [23, 24]. The genomic and
proteomic data bases (like those located in NCBI – the US National Centre
for Biotechnology Information) contain a big amount of information about
genes and proteins from numerous different species [25-27]. For example
the GtRNA database [27] contains more than 111,000 tRNA sequences in
which conserved motifs of length 22 are searched in the tRNA loops, D-
loop, anticodon-loop and Tψ-loop [28-30]. Because tRNA sequences pos-
sess about 70 bases, this search corresponds to about 2x66 105≈107 subse-
quences examination and comparison with reference sequences (Figure 5).
Figure 5 Result of a data mining in the NCBI database GtRNA [27] ex-
tracting 18196 occurrences of motifs related to tRNA loops among about
107 possibilities (the present list contains only the beginning of the results).

12 J. Demongeot, M. Jelassi, C.Taramasco
1.2.4. Compressing before visualizing
Before the deep learning phase, it is recommended to compress the data,
for improving the rapidity of the learning, but also for allowing an easy
visualization by the people manipulating the raw information and by the
data provider, that is the patient. The compression can be obtained by i)
reducing the dimensionality of the observables (e.g., using Principal Com-
ponent Analysis or Independent Component Analysis), ii) discretizing
(even binarizing) the data [31-33] (e.g., using fuzzy or Fourier transform,
which retain respectively only Boolean or few parameters values) before
the steps of classification [34-36] or data mining [35-37].
Apart the medical imaging [23, 24, 38], the main sources of medical in-
formation correspond to devices in which a cell, tissue, organ or the entire
body is located in a physical field (electromagnetic, gravitational, acoustic,
etc.) whose intensity varies abruptly, leaving the cell, tissue, organ or en-
tire body to return to its resting state, this phenomenon being called “relax-
ation”: it is the case for Nuclear Magnetic Resonance (NMR), mass or
NMR spectroscopy, Ultra-Sound (US) echography, electrophysiology, etc.
In the last case, signals like ECG (Electro-Cardio-Graphy), EEG (Electro-
Encephalo-Graphy) or EMG (Electro-Myo-Graphy) can be considered as
relaxation signals, and compressed using their decomposition on a base of
solutions of various pendulum differential equations.
The classical Fourier and wavelet transforms correspond for example to
the decomposition of a signal on a family of solutions of differential equa-
tions, namely the simple pendulum equation for Fourier transform and the
damped pendulum equation, for wavelet transform (Figure 5 bottom). The-
se transforms are well adapted to the case of quasi-symmetric and periodic
signals (i.e., whose wave shape has an internal symmetry on a hemi-
period, like the sine function), such as the respiratory rhythm (Figure 4
top), for which with few Fourier harmonics, the original shape can be re-
constructed with a small controlled error. It is the same for damped quasi-
symmetric signals, such as the vibration of the basilar membrane after a
short sound stimulus, which is well compressed by wavelet transform.
But relaxation signal like ECG needs a large number N of Fourier or
wavelet harmonics to be well compressed (e.g., N=20 for ECG and Fourier
with Signal-to-Noise Ratio (SNR) of about 20 dB using online site http:
//lpsa.swarthmore.edu/Fourier/Series/WhyFS.html#Electrocardiogram_).
In these cases, a smart compression method consists in decomposing the
signal on a family of approximate polynomial solutions (called dynalets

13
[39, 40]) of the anharmonic pendulum equation, proposed first in its ana-
logic version (the “singing arc lamp”) in 1900 by W. Du Bois Duddell,
then mathematically studied by H. Poincaré in 1908 (who invented first for
this equation the term limit-cycle for its unique attractor), and eventually,
proposed by B. van der Pol in 1926 to model the heart functioning. The
Figure 5 top shows the decomposition of the respiratory wave into one
fundamental and three Fourier sub-harmonics, Figure 5 bottom shows the
pendulum equations used in the three decomposition methods, Fourier,
wavelets and dynalets, and Figure 6 proposes an example of dynalet de-
composition of the cardiac ECG signal.
Figure 5 Top: Fourier decomposition of the respiratory wave into one fun-
damental and three Fourier sub-harmonics, giving a correct reconstitution
and allowing for a graphic representation (called respiratory “aster”) in the
amplitude-phase plane of the signal. Bottom: the 3 pendulum equations
corresponding to the 3 transforms, Fourier, wavelets and dynalets [39, 40].
Fourier wavelets dynalets

14 J. Demongeot, M. Jelassi, C.Taramasco
Figure 6 Top: initial position in the phase plane x0y of the initial van der Pol limit
cycle (in green clear) and ECG signal (in red), with the final fit between van der
Pol (in dark green) and ECG signal. Bottom a): fundamental dynalet component
extraction from the original experimental ECG signal (in red). Bottom b): match
between the experimental ECG (EXP in blue) with the sum S (in violet) of the fun-
damental dynalet X1 (in red) and the first sub-harmonic (X2 in green), with a trans-
formation in the (xOy) phase plane consisting in translating/scaling x and y axes, in
order to obtain the best fit for a cost function based on the Hausdorff-distance be-
tween 100 sampled empirical mean points and the set of points of same phase ex-
tracted from the van der Pol limit-cycle [39, 40]. Bottom c): calculation of the se-
cond sub-harmonics by subtracting the fundamental plus the first harmonic
component (S in violet) from the sampled original ECG signal (EXP in blue).
a)
EXP
X1
X2
S

15
Another efficient way for compressing the signal is to extract a minimal
information corresponding only to the succession of the sense of variations
(called monotony signature): +, _ or = of the signal, this information being poorer
than that coming from rank statistics: observing the rank, we can reconstruct the
signature, but inverse is false [41]. Retaining from the signal its monotony
signature, consists only in building the succession of + (if intensity of the signal
increases), - (if it decreases) or = (if it remains constant) on a succession of
windows along the time interval of recording. For example, on Figure 7, the
monotony signature of the green signal (equal to the number of entrance in
bedroom of a person at home (suffering Alzheimer’s disease with an important
cognitive impairment related to a loss of the short-term memory) during a day
equals: + - + - + - + - + - + + - - + - - - - + + + -, signing the probable existence of
a compulsive activity such as a pathologic perseveration (often observed in
neuro-degenerative diseases).
Figure 7 Evolution during successive 25 days of the number of actimetric
events of entrance during the nychthemeron (day/night 24h interval) in dif-
ferent rooms at home. The orange circles emphasize discrepancies between
the monotony of the green signal (entrance in bedroom) and the black one
(entrance in living).

16 J. Demongeot, M. Jelassi, C.Taramasco
A simple statistical test of equality of binomial variables shows that
there is no significant difference (p=0.05) between the monotony signature
of the number of entrance in living (black signal) and of entrance in bed-
room (green signal), proving the existence of a correlation between these
two signals of entrance. The visualization of the intervals of monotony of
various signals (as on Figure 7) by indicating their sense of variation al-
lows for identifying rapidly the discrepancies between signals, and hence,
permits to trigger a rapid alarm concerning a pathologic behavior from the
personalized actimetry recorded at home.
1.2.4. Clustering before interpreting
One of the goals of the big data techniques is to classify, i.e., to dispatch
the observations into a family of clusters, in which the mean value of the
cluster represents a statistical individual (in general virtual) representative
of the cluster population, if it is homogeneous (case of an unimodal distri-
bution of the clustervalues, corresponding in general to a quasi-Gaussian
shape for the distribution of the observed variables) and represents nothing
if the distribution is heterogeneous (multi-modal). If the signal has been
sufficiently compressed, the phase of clustering is rapid by using a classi-
cal unsupervised tool like k-means (introduced by J. MacQueen in 1967).
Similar methods like the “nuées dynamiques” by E. Diday (1971) and self-
organized maps by T. Kohonen (1981) are also easy to use in case of a
wide collection of data to classify, and, if these methods do not give results
easy to be interpreted, the support vector machine (SVM) by V.N. Vapnik
(1991) has the advantage to give a probabilistic explanation of the obtained
clustering. The final step of the deep learning can use one or more classifi-
cation methods (e.g., a supervised method) following the above approach-
es. The Figure 8 shows a classification of respiration shapes (after normal-
ization of their period) based on other ideas coming from probabilistic
algorithms: Hidden Markov Models (HMM) by R.L. Stratonovich (1960),
Expectation Maximization (EM) by A.P. Dempster (1977), Final Markov
Classes (FMC) by B. van Cutsem (applied in 1984 [42]), called now Mar-
kov CLuster (MCL) by S. van Dongen (2000), and Classification Expecta-
tion Maximization (CEM) by G. Celeux (1991). The classification of the
Figure 8 has been obtained using the FMC classification algorithm [42]
and the obtained classes have been proved to be robust after reapplication
of the same method on same individuals 4 years after the first recording
[43]. The association of a FFT on the respiratory signal and of a Markovi-
an clustering algorithm has permitted to treat rapidly 1000 patients and to
show the existence of a respiratory “personality”, attested by the perma-

17
nence of their respiratory profile, even if occur intercurrent diseases or
change of habits of life.
Figure 8 Top: respiratory waves of one class obtained by the final classes
algorithm [42]. Bottom: respiratory waves of the same class obtained from
records made on the same subjects, but 4 years after [43]. The wave shapes
derived from harmonic analysis of the digitised airflow signal (approxi-
mately 50 breaths per subject), and have been plotted in normalised time.

18 J. Demongeot, M. Jelassi, C.Taramasco
Another method of clustering available for big data sample comes from
the functional estimation. For example, one device providing spectral
signals from protein involved in cancerogenesis is the MALDI-TOF mass
spectroscope. Dataset originated from studies of colorectal cancer are
available in [44] and correspond to diseased subjects (Figure 9 Top) and
subjects of control (Figure 9 Middle) so that the data set contains 112
spectra of length 16331, among which 64 spectra are those of diseased
patients. This particular study contains only a few part of the 250,000
possible protein spectra (20,000 for proteins expressed by the human
coding genome and 230,000 for immunoproteins), but the clustering
method is the same for more massive cancer data studies.
After denoising using wavelets the mass spectroscopy signals, the use
of functional Principal Components Analysis (PCA) allows for estimating
the local functional likelihood of the curves after randomly partitioning
these curves into a learning set of size 80 and a test set of size 32. Based
on the optimality criterion for reducing the classification error rate, the
number of main components used for diminishing the dimension of the da-
ta and constructing a semi-metric permitting the classification, is equal to
four, giving an average error rate of 0.06 ± 0.056* for 100 test specimens
[45, 46], which corresponds to a discrimination between the mass spectra
better than one made with the classical kernel methods of estimation (aver-
age error rate of 0.072 ± 0.033) [47].
The Figure 9 Bottom shows the performance of the functional Principal
Components Analysis (PCA) and local functional likelihood approach [46]
in terms of Signal-to-Noise Ratio (SNR) compared to the functional Partial
Least Squares (PLS) regression [47], showing a disjunction of the spectra
between the two classes (cancer and control), which remain confounded in
the PLS method. This approach is available for huge data sets and for a
large number of classes, offering then an alternative to the traditional esti-
mation methods, which consider the minimization of the least absolute rel-
ative error for regression models.
The strong and the uniform consistencies of the constructed estimator
as well as the mean squared convergence rate and the asymptotic normality
of the proposed estimator are proved in [45, 46], showing that it is possible
to combine quantitative methods of functional non-parametric estimation
(i.e., with a minimum of hypotheses on the probabilistic structure of the
data) with classical qualitative techniques of non-inferential data analysis,
such as the functional Principal Component Analysis.

19
Figure 9 Top and Middle: MALDI-TOF mass spectra from the group G1
of patients suffering colorectal cancer (top) and control group G2 (middle)
recorded on the m/z interval of 900-11160 Da. Bottom: Signal-to-Noise
Ratio (dB) compared between the wavelet-PCA and wavelet-PLS methods.
SNR (dB)

20 J. Demongeot, M. Jelassi, C.Taramasco
1.3 The medical wireless sensing
The World Health Organization (WHO) has codified health disability in
the International Classification of Functioning (ICF), which provides a
standard for the description of the information concerning the well func-
tioning or the disability of a person [48]. The IoT approach in e-Health
gives the possibility to build suitable platforms to realize ubiquitous health
studies using body and/or environmental sensors and to upload the record-
ed data to servers to be stored, treated and interpreted for triggering alarms
at individual or population levels. Mobile Health (m-Health) emerged also
recently in healthcare, based on smart phones using several facilities like
Bluetooth for interfacing sensors measuring physiological parameters in
interoperable environments, such as home monitoring systems for aged-
care [49-60], appeared also recently, based not only on sensors and actua-
tors, but also on wearable, implantable or microcapsule devices connected
through wired or wireless networks to a service center with diagnosis and
therapeutic facilities.
These devices can assess images, body motions, sounds and ambient
parameters (light, temperature, humidity, etc.), vital signs (blood pressure,
respiration rhythm, body temperature, heart/pulse rate, body weight/fat,
blood oxygenation, ECG, etc.), sleep patterns and other health parameters
related to daily activities as well as to social interactions. All these new e-
Health, IoT and m-Health devices are participating to the “tsunami” of
medical data, to the processing of which the big data techniques combined
with recent statistical tools [61-64] seem particularly well adapted.
In the e-Health systems, physical sensors, for example, register the po-
sition and the movements of the patient and of their caregivers at home,
capturing information about changes of many external physical fields,
modified during daily activities by the elderly: thermal field (for early de-
tection of thermal discomfort, premise, in the worst case, of a move toward
malignant hyperthermia) using smart vest and bracelets, gravitational field
(detecting for example abnormal accelerations of the trunk) using smart
vests, bright field using infrared sensors and surveillance cameras, elec-
tromagnetic field (used for presence detection) using magnetometers and
acoustic field (allowing localization and identification) using microphones,
possibly combined for detecting the size of one step, or detecting a fall by
using accelerometers or microphones.

21
e-Health smart home technologies can be classified according to their
function [65]:
• Physiological monitoring (measurement of vital signs),
• Functional monitoring (measurement of the general activity level,
meal intake, motion, etc.) and emergency detection (abnormal or
critical situations such as falls),
• Safety monitoring (occurrence of environmental hazards such fire
or gas leak) and assistance (such as automatic turning on/off path
to bathroom lights when getting out of the bed),
• Security monitoring and assistance to detect and manage human
threats such as intruders
• Social interaction monitoring and assistance like phone calls, vis-
its, coaching and participation in social activities. Assistance in-
cludes technologies allowing virtual participation in group activi-
ties, video-calls with family and friends, etc.
• Cognitive and sensor assistance like automated or self-initiated
reminders (medication reminders, lost key locators, etc.) and task
instruction technologies.
Devices used in healthcare, especially those based on IoT and m-Health
technology, could help not only in the care of elderly people but also in the
management of medical surveillance of chronic diseases such as cardio-
vascular, metabolic, pulmonary, renal or neuro-degenerative diseases. For
type 2 diabetics for example, sensors allow for following two potential
complications of diabetes: nephropathy and diabetic foot. Several devices
have been cited in the review [66] such as implantable glucose sensor, en-
do-radio-probe, radio pills, etc. Implantable glucose sensors have been
created based on the ‘enzyme electrode’ principle, in order to get the glu-
cose concentration [67]. Endo-radio-probe [68], a microelectronic device
introduced into the body to record physiological data not otherwise obtain-
able, has been used to develop an in vivo drug delivery. The principle is to
swallow a small transmitter (possibly coupled with a micro-camera) in or-
der to monitor digestive tract parameters such as pressure, temperature,
and pH. The system consists of a controller, a radio frequency (RF) trans-
mitter and a receiver [69].
1.3.1. Smart homes and actimetry
Actimetry sensors record positions and movements of a patient and his
caregivers at home (Figure 10), capturing information about external phys-
ical fields changes (thermal, gravitational, light, electro-magnetic and
acoustic) during the daily activities of a person [70-74].

22 J. Demongeot, M. Jelassi, C.Taramasco
Figure 10 Left: in the living room, pressure sensors under the feet of furni-
ture and web of pressure on the chairs, armchairs, sofas and beds (gray);
lighting sensor (yellow) and infrared presence sensor (red). Middle: in the
kitchen, pressure sensor under a bottle containing a liquid (gray), water
flow sensor at the tap (blue) and electronic switches for opening / closing
doors (green). Right: in the toilet, water level sensor (blue) and chemo-
sensor or resistivity sensor for urea concentration (green).
1.3.2. Thermal sensors
Early detection of thermal discomfort involves capturing multiple infor-
mation concerning both the temperature and the resistivity of the skin, the
humidity of the room and the intensity of ambient air flow. The develop-
ment of micro-power acquisition and processing devices devoted to these
data [71] allows for considering the possibility to prevent early abnormali-
ties of thermoregulation caused by a heat wave. The acceptability of sen-
sors allows for capturing the surface temperature on the chest, wrist and
ankle, through the assimilation of these sensors to familiar objects within
usual clothes, watches or lockets.
• Pyrosensor away at body temperature
The medicalized antimicrobial lighting sconces by the company Legrand®
bring innovative techniques in terms of decision, connectors, and light
source. They incorporates a magnetic plug (type Apple®, with which they
share the patent, but bigger), a manipulator, standard jacks (2 + earth, for
electricity) and FTP (Foiled Twisted Pair) jacks for medical computer
equipment [72]. The temperature measurement sensor is a thermopile in-

23
cluded in a fixed pyrosensor placed on the light applied to the top of the
bed at the living place, allowing for a double monitoring of presence and
hyperthermia.
• Sensor thermal drop
The sensor works by directly measuring thermal infrared radiation in the
human environment. It can determine the presence of one or more persons
in a room. This information is sent to a microcontroller that transmits data
(4x4 matrix with 8 temperature levels, see Figure 11) via a USB serial
communication to a PC, running the algorithms needed to detect the falling
speed [73].
Figure 11 Left: thermal profile of the person. Middle: thermal sensor.
Right: 4x4 matrix, giving the profile summary with 8 temperature levels.
1.3.5. Gravitational sensors
• Fall sensors
A sensorized vest including a fall sensor for detecting abnormal accelera-
tion of the trunk has been developed and patented [60]. This information,
cross-checked with actimetric data (pressure sensors on the ground, radar
surveillance, infrared sensors, magnetic sensors of door opening, etc.) can
trigger a selective information to the victim of the fall and an emergency

24 J. Demongeot, M. Jelassi, C.Taramasco
intervention if he does not want or cannot react (at least verbally). The ac-
ceptability of the fall sensor is strongly linked to its specificity (absence of
false positive ensured by the redundancy with environmental sensors).
• Pressure sensors
The principle of the pressure pad is the same as the anti-decubitus actimet-
ric mattresses. In collaboration with the company Léas®, a prototype has
been developed, which allows for measuring in real time the pressure on
the buttocks of a subject using a wheelchair [75]. This cushion is made of
two right and left hemi-cushions (Figure 11). Each half-cushion contains
6x12 pressure sensors. Each pressure sensor has an area of approximately
1 cm2. The distribution of the sensors in the pad is non-linear, to allow bet-
ter resolution in the ischiatic zones, where the risk of bedsores are higher.
The sensors operate through a semiconductor powder evenly distributed in
a polymer shell. This powder has elastic properties and acts on the princi-
ple of percolation: when its volume changes under the effect of a pressure
force, its conductivity increases and the variation of current passing
through the powder is thus a function of the pressure exerted. Each sensor
is connected to an electronic system, which enables measurement of an en-
coded 4-bit electrical potential (16 pressure levels).
The pad is coupled with another prototype, a lingual stimulator named
Tongue Display Unit® (TDU), allowing the feedback to the subject [76].
After the seminal work by Kazimierz Noiszewski [77-81], who developed
the first device of substitutive vision by tactile stimulation, called El-
ektroftalm® or “artificial eye” (1897), followed by the Optophone® of
Fournier d'Albe (1912) [82], devices have been developed and validated by
Samsó Diaz first (1962) and then, by Bach-y-Rita (1969) in a population of
deaf animals and blind humans, for which acoustic or visual information
captured by microphones or video cameras was transcoded in electrotactile
stimulations of skin or tongue [83-85].
The tool has been adapted (including wireless) to people at risk of bed-
sores. The subject can keep the electrodes in contact with his tongue,
mouth closed. Saliva having a good conductivity, TDU only requires a
voltage of 5 to 15V and a current of 0.4 to 4mA to stimulate lingual recep-
tors. When an electrode is activated, the subject feels a "tingling" on the
surface of his tongue. A subject without spinal cord injury can thus per-
ceive and interpret information provided by the TDU electrodes placed on
his tongue and adopt a postural attitude adapted to the information coded

25
in false color (Figure 11) in order to limit the overpressure zones. The ac-
ceptability of the transmission of the dermal pressure information requires
the incorporation of stimulation electrodes in an artificial palate, which
may be that of the dental prosthesis.
The pressure information can also be delivered as a vibration signal at
the posterior surface of the upper incisors causing a correction of posture,
which can be unconscious in healthy subjects, like in the case of the der-
mal pressure messages from the sole of the feet [86, 87]. The noninvasive
nature of the capture and restitution has been demonstrated, and further
studies could show that after a learning phase, the recruitment of the corti-
cal areas of projection of the tongue sensitivity, could definitely prove the
phenomenon of substitution and habituation by the stimulated subject
(process similar to the gradual habituation of the tongue to a new mouth
environment, even slightly modified, after dental work by a dentist).
Figure 11 Top left: the signal on the watch indicates the left side, opposite
to where the pressure is the greatest. Top middle: discretization of the in-
formation of pressure. Top right: same information given on a smartphone.
Bottom: pressure recorder cushion.

26 J. Demongeot, M. Jelassi, C.Taramasco
1.3.6. The multi-sensor fusion and alarm triggering
• Generalities on the multi-sensor fusion
Alarms generated by a monitoring system at the living place (e.g., at home
or in a residency for senior) do not directly concern the diagnosis of a dis-
ease, as could a medical expert system, giving a code related to the Inter-
national Classification of Diseases (ICD / DCI) of the World Health Or-
ganization (WHO) [48], but rather indicate a functional deficit, as codified
in the International Classification of Functioning, Disability and Health
(ICF / ICF) of WHO [48], in order to provide a standard in the description
of dysfunctions and disabilities, which belong to an ontology of common
concepts that concerns:
- organic functions and anatomical structures of individuals
- activities of individuals in areas of social life in which they participate
- personal and environmental factors that influence this participation.
ICF does not classify people, but describes them through multiple hier-
archical as well as non-hierarchical classes, which themselves consist of
categories. Each category is coded qualitatively in order to describe the
sensory, motor and cognitive limitations, restrictions to a social participa-
tion, environmental barriers, etc. ICF information can trigger an alarm, if
the person enters a new set of critical classes, which require attention (im-
mediate or possibly delayed).
• Multi-sensor integration: the example of detecting the fall
The function of the fall surveillance system is to detect a person who loses
his balance and falls on the floor, compromising his safety and physical
and / or intellectual integrity. The operating principle is based on a sound
or an infrared signal passing over a critical threshold and generating an
alert. The infrared signal comes from a sensor capable of delivering a local
temperature pattern in a 4x4 pixel matrix, within a sensing range of about
9 m2. This information is sent to a microcontroller that transmits data via a
USB serial communication to a computer of the ODROID family, integrat-
ing these data with the sound signals from a microphone or those of a pres-
sure pad, and executing the algorithms required for detecting the fall as a
critical negative acceleration, the common feature of fall sensors [88-94].
The system can be installed for example in the bedroom, the bathroom
and the toilet, frequent places of nocturnal fall. After an adjusting phase,
the system automatically sends a SMS or email alert to the medico-social

27
environment of the monitored person. The fusion algorithm combines in-
formation from infrared sensors with that coming from CMT (Microphone
Technology Coincidence) microphones and pressure sensors [94]. The
specific constraints of the system are:
- the measured precision is very different for different sensors: a room
or a large area of the room for infrared sensors, 15° azimuth for CMT mi-
crophones and at most 1/2 m2 for pressure sensitive tiles or pads,
- the presence information from different sensors is not always guaran-
teed: CMT will be effective only in case of fall with noise, cry or call by
the monitored person (in this case, CMT localization signal has to be fil-
tered by the information coming from an associated smart sound system,
which detects and interprets the speech or the abnormal silence)
- the detection has to be performed in real time and to be reliable even
the person is not alone.
• Example of aid to the determination of the plantar balance
A podiatrist wishing to assess the sheet balance before care and massage
and in case of feet ulcers in a complicated type 2 diabetic, can use an im-
portant information on how abnormal is the walk of his patient, if it tends
to antalgic gait, a limp adopted so as to avoid pain on weight-bearing
structures, for example an equinus gait, characterized by a tiptoe walking,
to escape a pain of a heel ulcer.
A smart sock (respectively sole), developed by the company Tex-
isense® [86] (respectively FeetMe® [87]) tracks the patient’s movements
when walking and a bone imaging (Figure 12), recorded by a portable ul-
trasonic device (e.g. from GE Healthcare®), serves to mapping a default of
ossification of the calcaneus (bone whose resorption/reconstruction by os-
teoclasts and osteoblasts has a turnover/remodeling phase of about 17
months), when it is facing a heel ulcer. The bone tissue is indeed constant-
ly renewed, process disrupted in diabetes, by a loss of osteocalcine secre-
tion and plantar pressure stimulation (due to the antalgic avoidance of the
ulcerated parts of the feet).
• A necessary data fusion
All the data collected from dedicated fall sensors and from environmental
actimetric sensors need to be analyzed separately and then, merged by us-
ing a fusion algorithm, in order to be sent to a big data classifier, able to
detect the entrance in a zone at risk. The Figure 13 summarizes the differ-
ent steps of the fusion procedure and alarm triggering. The scoring using

28 J. Demongeot, M. Jelassi, C.Taramasco
ICF classification gives a scale on which the assignment to a class of risk
is made. The alarm information is transmitted both to the patient (if he has
still the cognitive and sensory-motor capacity to interpret it and to react)
and to his caregivers, as well as to his medico-social environment, trigger-
ing for example an immediate help at home. An effective fusion of physio-
logical variables coming from body sensors and environmental variables -
more reliable and accurate than the natural one's – coming from external
sensors (on walls or furniture) is for example crucial to enable individuals
with spinal cord injuries or with somatosensory loss in the feet (e.g., from
a diabetic peripheral neuropathy) to become aware of a localized excess of
pressure at the skin/seat interface and/or postural orientation and thus to
make adaptive postural corrections to prevent the formation of pressure ul-
cers and/or falls.
Figure 12 The area of antalgic gait on the forefoot to escape ulcerative
heel pain is recorded by the smart sock of Texisense®, which captures the
pressure map under the feet (area in red on thumbnail image).
Calcaneus
Ulcerated zone
Pressure zone

29
Figure 13 Top: fusion of data from different types of exo- and endo-
sensors detecting gravitational, acoustic, optical, thermal and electromag-
netic fields. Bottom: fusion of data from actimetric sensors and presence
and time information sources needed to calculate a profile from IFC scores
and trigger an alarm in case of life-threatening risk, like a fall.

30 J. Demongeot, M. Jelassi, C.Taramasco
1.3. 8. Critical profiles and alarms: visualization, analysis
and interpretation
Let us consider now the central vegetative system ruling the cardio-
respiratory activity. Its functioning can be summarized on Figure 14. The
central vegetative system has two components; (i) the bulbar respiratory
center with its inspiratory (I) and expiratory (E) neurons, the first ones ex-
citing the second ones and inversely the second ones inhibiting the first
ones, and the cardio-moderator (C). These components rule the main pe-
ripheral actuators of the cardio-respiratory system, i.e., the diaphragm and
the heart controlled by the sinusal node (S). In the case of dysfunction of
the cardio-respiratory system, acute (like a cardiac failure) or chronic (like
a pathologic respiration in an obese person), the patient and his caregiver
and medico-social environment have to be rapidly informed to correct the
dysfunction and avoid future complications. Several devices like the smart
clothing Visuresp® [95] can record in real time both the respiratory and
the cardiac rhythm and send the information to a center monitoring many
patients like the Kaplan center of cardio-respiratory rehabilitation at Val-
paraiso in which an innovative remote rehabilitation program for the car-
diovascular patients of Dr. Gustavo Fricke Hospital has been developed by
the University of Valparaiso and the Viña del Mar Quillota Healthcare
Services in collaboration with the Jorge Kaplan Foundation, thanks to an
award of the Innovation Fund for Competitiveness (FIC) of the Regional
Government of Valparaiso [96].
Figure 14 The central vegetative system made of the bulbar respiratory
center with inspiratory (I) and expiratory (E) neurons and of the cardio-
moderator (C), ruling the main peripheral actuators, namely the diaphragm
(not represented) and the heart controlled by the sinusal node (S). The var-
iables x, y, w and z represent respectively the electrical activity of the cen-
ters E, I, C and S.

31
In order to visualize, analyze and interpret rapidly the data coming from
the Visuresp® sensors, it is necessary to enter in the last phase of the big
data approach, the mathematical modeling, using here non-linear differen-
tial equations to account for the relaxation behavior mainly of the cardiac,
but also of the respiratory components, for which the modeling with the
van der Pol system has been chosen [40].
The van Pol system representing the rythmic respiratory activity reads:
dx/dt = y, dy/dt = -x+ε(1-x2)y, (1)
where ε represents the anharmonic parameter of the oscillator, with a free
run (or proper period) τ equal (near the bifurcation of the van der Pol limit
cycle obtained for ε = 0), to the ratio τ = 2π/i, where i = (2-ε2/2)1/2 is the
imaginary part of the eigenvalues of the Jacobian matrix of the system (1):
0 1
J =
-1 – 2εxy ε(1-x2)
The van der Pol system representing the rythmic cardiac activity reads:
dz/dt = w, dw/dt = -z+η (1-z2)w + k(y)y, (2)
where η is the anharmonic parameter and k(y) is the coupling intensity be-
tween I and CM. The entrained period of the cardiac oscillator is equal to:
T = 2π/(2-η2(1-(k(y)y)2)2/2)1/2.
The values of ε et η are fixed by the proper periods of the respiratory
(4s) and cardiac (1s) oscillators, then k(y) can be obtained by measuring
the instantaneous cardiac period T (which is just the inter-beats duration)
and by calculating the slope of the regression line between T and the res-
piratory activity (represented by the actual inspiratory time τ where occurs
the cardiac beat of period T). This slope is directly related to the correla-
tion coefficient between T and τ (cf. Figure 15 showing the periodical evo-
lution of T, which proves the coupling between the two oscillators). The
integrity of this coupling allows the bulbar vegetative system for adapting

32 J. Demongeot, M. Jelassi, C.Taramasco
its electrical activity to the effort: first the breathing is ruled by the will or
entrained by a muscular activity and secondarily it entrains the heart. Such
a capacity of adaptation disappears in degenerative diseases like the Par-
kinson or the diabetes. Watching a parameter like ρ is then particularly in-
teresting in the elderly people and the Visuresp® and environmental acti-
metric sensors permit the emergence and the management of a crucial
knowledge about the vegetative regulation of the cardio-respiratory sys-
tem.
By observing large samples of patients, it is possible to constitute
standards of actimetric or vegetative values and then, to extract a physio-
logical knowledge from these data, e.g., that concerning the integration in
the vegetative system of the respiratory and cardiac controls. By analyzing
the nycthemeral (day-night) curves of a patient equipped by a device Vi-
suresp® and/or other cardio-respiratory specific sensors, it is observed that
the cardiac instantaneous period (just the lapse of time between two cardi-
ac beats) is anti-correlated with the inspiration duration as well as with the
time at which the inspiration occurs in the cardiac cycle (Figure 15): the
cardiac rhythm is accelerating during the inspiration, and the heart deceler-
ates during the expiration.
The progressive disappearance of this coupling during neurodegenera-
tive pathologies like Parkinson’s or Alzheimer’s diseases allows for diag-
nosing early the entrance in the chronic dysfunctions of the cardio-
respiratory system related to these diseases. In order to restore an healthy
functioning, for example in a patient suffering a broncho-obstructive pa-
thology, a biofeedback rehabilitation is possible, which would advice the
patient to follow on a screen an ideal respiratory rhythm calculated from
his respiratory signature [43], compared to the normal cardiorespiratory
behavior of the class to which he belongs, class provided by the big data
procedure of clustering.
The assignment to a class of patients at risk can be made also by con-
sidering the social network surrounding the patient, in which there are
normal people to imitate (in particular in their alimentation or way of life
habits), and pathologic patients to avoid, especially those having a nega-
tive influence on their human environment (Figure 16). Depending on the
state of a patient and also on his environment in the social network to
which he belongs, it is possible to personalize a preventive and therapeutic
education program [97], whose aim is to globally reduce the occurrence of
diseases in the social network. This point will be treated in the Discussion.

33
Figure 15 Top: representative sections of the individual original airflow
recordings from two studies separated by 4 years, with their physical char-
acteristics at the time of the first study, showing the conservation of the
respiratory profile, then defining the signature of the respiratory personali-
ty [43]. Middle: evolution of the instantaneous cardiac period, anti-
correlated with the inspiration duration. Bottom: biofeedback rehabilitation
using the smart shirt VisuResp® of a patient suffering from a respiratory
disease, such as an asthma.
cardiac period
ECG
Air flow
inspiration

34 J. Demongeot, M. Jelassi, C.Taramasco
Figure 16 Top: Periodic breathing of obese people, with crescendo-
decrescendo alterations in respiratory effort and airflow without central
apneas, and normal flow rate of normal weight people without cardio-
respiratory pathology. Bottom left: social network with normal weight (in
blue) and overweight or obese (in red) people, with indication of the social
links (the surface of a node is proportional to the number of such links).
Bottom right: distribution of the number of friendship links over the nor-
mal (in green) and overweight/obese (in violet) populations, showing a
bimodality of the distribution into the last one.
1.4. Discussion
The complex system approach aims to extract emergent properties from
the observation and modelling of the entities in interaction, which provide
large spatio-temporal data sets corresponding in general to robust complex
systems well regulated in the framework of the homeostatsis, notion which
characterizes the stability of the genetic, metabolic and physiologic net-
works that regulate all the interactions inside a living system [98].
The big data approach would converge with the complex systems one,
because the explanatory modelling remains the ultimate step in the
Normal airflow in normal weight person
Pathologic airflow in obese people
Nb#nodes#
Normal
Overweight or obese
Normal
Overweight or obese

35
analysis of mass data, and the theory of complex systems makes it possible
to define a minima models that can account for the variability of data with
a small number of parameters and also to extract emerging properties,
serving as a refutation tool for making these models more realistic and
therefore more predictive, especially in medical applications. The future
strategy in medical data processing would be thus constituted by a
successive use of methods coming from big data tools and others related to
the theory of complex systems, combined in order to face all the
challenges posed by the mass information in health (especially in e-health),
from its visualization to its modelling [99].
An example of such a collaboration between the two paradigms, big
data and complex systems, is offered by the social networks. The complex
system approach of social networks allows for defining a new notion of
centrality, called the entropy centrality [91], taking into account the heter-
ogeneity of the states of the neighbors of a given node i in the network,
and not only the connectivity of its neighborhood in the interaction graph:
Centropy
i = - Σk=1,…,s
ν
kLog
ν
k,
where
ν
k denotes the kth frequency among the s frequencies of the histo-
gram of the state values observed in the neighborhood of i, denoted Vi, i.e.,
the set of the nodes linked to the node i.
This new notion of centrality is for example more useful than the previ-
ous ones to detect which obese patient is a good candidate to be educated
through personalized therapeutic advices given at home (depending on the
observation at home of his actimetric and metabolic data) in order to trans-
form his pathologic state in a normal state, because this candidate can in-
fluence efficiently a heterogeneous environment, pushing his neighbors to
recover their weight normality. An illustration of this fact is done in the
example of the Figure 17, which represents the result of the observation of
a large social networks with obese, overweight and normal weight persons
in familial or friendly relationships, having positive or negative influences
on their neighbors. This influence, if it comes from an obese and if it is
positive, can push his normal friends to adopt his bad way of life and nutri-
tion habits, hence causing their passage from the normal weight state to the
overweight one. In order to modify this bad influence in the social net-
work, it is possible to detect the critical obese people having in their envi-
ronment a sufficient number of normal or overweight people susceptible to
be influenced.

36 J. Demongeot, M. Jelassi, C.Taramasco
A big data approach in a huge social network is capable to detect rapidly
the critical “hubs”: there exist classical notions of centrality (degree, be-
tweenness, closeness and eigenvector centralities, well known in graph
theory) allowing for identifying the critical nodes susceptible to become
good target of a personalized therapeutic education, and the new notion of
entropy centrality has been tested with respect to the previous ones.
By comparing the results obtained thanks to a personalized education
leading obese nodes to recover the normal weight state, we observe in the
example of Figure 17 that after educating only the 21 individuals having
the most important entropy centrality (Figure 17 Bottom right), all the
population is going back to the normal weight state, but that needs 68 indi-
viduals with the degree centrality, and 85 individuals with the eigenvector
centrality (Figure 17 Top and Bottom left). Then, the best public health
policy against the obesity pandemics consists in using the notion of entro-
py centrality to select the targets of the therapeutic education.
The example of the Figure 17 shows the necessity for the big data and
complex system approaches to collaborate, because the big data techniques
are very useful in the first steps of data processing (compression, visualiza-
tion and interpretation), but have to be completed in the explanatory phase
of modeling, that can in general be described in terms of complex system,
for representing interactions between healthy persons and patients suffer-
ing a disease [100], and also for dealing with the disease pathogenesis, by
considering the network of all the genes, metabolites and environmental
factors involved in the occurrence of the disease [101]. The two
approaches are therefore not contradictory, but complementary, and must
demonstrate in the future their effective collaboration, necessary for the
resolution, for example, of the new medical challenges posed at the
population level, for example by the obesity pandemic and the return of
infectious epidemics [102].
1.5. Conclusion
The big data approach used to organize data sets from e-health multi-
sensor systems makes it possible to classify observed individuals into
normal behavior classes requiring no specific support and into risk classes
requiring either a chronic surveillance without intervention, or an
immediate care triggered by an alarm calibrated on a classification
resulting, for example, from deep learning techniques in the population
studied (which may be several million in size, for example for cardio-

37
respiratory or obesity pathologies presented as examples in this chapter).
Figure 17 Top: representation of the interaction graph of a part of a huge
social network containing obese (red), overweight (pink) and normal
weight (green) individuals in interaction (family members or friends). The
nodes size corresponds to the in-degree (left), eigenvector (middle) and to-
tal degree (right) centralities. Bottom: threshold for the success of a thera-
peutic education of the N individuals having the largest entropy centrality:
after stabilizing the contagion dynamics, all individuals are overweight
(left) if N=20, and all individuals are normal (right) if N=21, this last
number being the critical threshold ensuring the success of the therapeutic
education.
Considerable progress remains to be made in the availability and
interpretation of data, especially for the patient and his helpers at his / her
place of life. The rapid and informative restitution of a summary of patient

38 J. Demongeot, M. Jelassi, C.Taramasco
data is more than an interesting scientific challenge, but also a legal
obligation in certain countries (such as France), an ethical necessity and an
effective means of empowering people, then more inclined to follow the
preventive and therapeutic recommendations made by the health services
(doctors and all para-medical professionals) concerning their illness and its
possible complications.
Acknowledgement. We acknowledge the Project PHC Maghreb SCIM
(Systèmes Complexes et Ingénierie Médicale) for its financial support.
References
1. Rhazes M (about 930). Manfe’ al Aghzie va Mazareha (Benefits of Food and
its Harmfulness). Tehran: Taban Press (1964).
2. Rhazes, M (about 925). De regimine sanitatis, or Epistula Alexandro de dieta
servanda. Juan de Sevilla Hispanus (Transl.) (about 1145). Paris: Antoine
Vérard (1497).
3. Bacon R (1280). Secretum secretorum cum Glossis et Notulis. In R. Steele
(Ed.). Opera hactenus inedita Rogeri Baconsi. Oxford: Clarendon Press
(1920), pp. 287-313.
4. Albæripæ W (about 1160). Tractatus Numerorum a ternario usque ad duode-
narium, ad Thomam monachum. Troyes: Médiathèque de Troyes incunable
969, pp. 195-196.
5. Shaiykh Bahā’ī, M (about 1610). Kholasat al-Hesab. The summa of arithme-
tic. In G.H.F. Nesselmann (Transl.). Essenz der Rechenkunst von Moham-
med Beha-eddi ben al-Hosain, al-Khulasa fil hisab. Berlin : G. Reimer.
(1843).
6. Tutte WT (1948). The Dissection of Equilateral Triangles into Equilateral
Triangles. Proc. Cambridge Phil. Soc., 44, 463-482.
7. Drápal A & Hämäläinen C (2010). An enumeration of equilateral triangle
dissections. Discrete Applied Mathematics, 158, 1479-1495.
8. Buteo J (1554). Opera Geometrica. Lyon: T. Bertellus.
9. Duchon J (1976). Fonctions spline et espérances conditionnelles de champs
gaussiens. Annales Scientifiques de l’Université de Clermont-Ferrand, 14,
19-27.
10. Duchon J (1976). Interpolation de fonctions de deux variables suivant le
principe de la flexion des plaques minces. RAIRO (Analyse Numérique), 10,
5-12.
11. Duchon J (1977). Splines minimizing rotation-invariant. Lect. Notes in
Maths, 571, 85-100.

39
12. Bernoulli D (1760). Essai d'une nouvelle analyse de la mortalité causée par la
petite vérole, et des avantages de l'inoculation pour la prévenir. Paris: Acad.
Roy. Sci.
13. d'Alembert J (1761). Onzième Mémoire : sur l’application du calcul des
probabilités à l’inoculation de la petite vérole ; notes sur le mémoire pré-
cédent ; théorie mathématique de l’inoculation In Opuscules mathématiques.
Paris: David.
14. Demongeot J, Fouquet Y, Tayyab M & Vuillerme N (2009). Understanding
Physiological & Degenerative Natural Vision Mechanisms to Define Robust
Contrast and Segmentation Operators. PLoS ONE, 4, e6010.
15. Demongeot J, Mattes J (1995). Neural networks and contrast enhancement in
medical imaging. In: Fogelman-Soulié F, Gallinari P, eds. ICANN’95, Paris:
EC2, pp 41-48.
16. Herve T, Dolmazon JM & Demongeot J (1990). Random field and neural in-
formation: a new representation for multi-neuronal activity. Proc. Natl. Acad.
Sci. USA, 87, 806-810.
17. Demongeot J & Fricot J (1986). Random fields and renewal potentials.
Springer Series, NATO ASI Serie F, 20, 71-84.
18. LeCun Y (1986). Learning process in an asymmetric threshold network.
Springer Series, NATO ASI Serie F, 20, 233-240.
19. LeCun Y, Bengio Y & Hinton G (2015). Deep learning. Nature, 521, 436-
444.
20. Zhao Y, Gao S, Gallinari P & Guo J (2015). Knowledge base completion by
learning pairwise-interaction differentiated embeddings. Data Min. Knowl.
Discov., 29, 1486-1504.
21. Dey N, Ashour AS & Bhatt C (2017). Internet of Things Driven Connected
Healthcare. In: Bhatt C, Dey N, Ashour A, eds. Internet of Things and Big
Data Technologies for Next Generation Healthcare. Studies in Big Data, 23.
Cham: Springer, pp. 13-33.
22. Elhayatmy G, Dey N & Ashour AS (2018). Internet of Things Based Wire-
less Body Area Network in Healthcare. In: Dey N, Hassanien A, Bhatt C,
Ashour A, Satapathy S, eds. Internet of Things and Big Data Analytics To-
ward Next-Generation Intelligence. Studies in Big Data, 30. Cham: Springer,
pp. 3-20.
23. Dey N, Ben Abdessalem Karâa W, Chakraborty S, Banerjee S, Salem MAM
& Azar AT (2015). Image mining framework and techniques: a review. Int.
J. Image Mining, 1, 45-65.
24. Ben Abdessalem Karâa & Dey N (2015). Biomedical Image Analysis and
Mining Techniques for Improved Health Outcomes. Hershey: IGI Global.
25. Dash S & Subudhi B (2016). Handbook of Research on Computational Intel-
ligence Applications in Bioinformatics. Hershey: IGI Global.
26. Lytras MD & Papadopoulou P (2017). Applying Big Data Analytics in Bio-
informatics and Medicine. Hershey: IGI Global.
27. GtRNA (2017). http://gtrnadb.ucsc.edu/

40 J. Demongeot, M. Jelassi, C.Taramasco
28. Demongeot J & H. Hazgui H (2016). The Poitiers school of mathematical
and theoretical biology: Besson-Gavaudan-Schützenberger’s conjectures on
genetic code and RNA structures. Acta Biotheoretica, 64, 403-426.
29. Demongeot J (2017). Proto-ribosome: a theoretical approach based on RNA
relics. BioRxiv preprint, doi: http://dx.doi.org/10.1101/182055
30. Demongeot J (2017). “Proto-ribosome” as new game of life. BioRxiv pre-
print, doi: http://dx.doi.org/10.1101/182055
31. Ramírez-Gallego S, García S, Mouriño-Talín H, Martínez-Rego D, Bolón-
Canedo V, Alonso-Betanzos A, Benítez JM & Herrera F (2016). Data discre-
tization: taxonomy and big data challenge. WIREs Data Mining Knowl. Dis-
cov., 6, 5-21.
32. Boukharrou RR, Ilié JM & Saidouni DE (2017). Time and Space Reasoning
for Ambient Systems. Int. J. Ambient Comput. Intell., 8, 38-57.
33. Mhetre NA, Deshpande AV & Mahalle PN (2016). Trust Management Mo-
del based on Fuzzy Approach for Ubiquitous Computing. Int. J. Ambient
Comput. Intell., 7, 33-46.
34. Dey N & Ashour AS (2016). Classification and Clustering in Biomedical Si-
gnal Processing. Hershey: IGI Global.
35. Acharjya D & Anitha A (2017). A Comparative Study of Statistical and
Rough Computing Models in Predictive Data Analysis Int. J. Ambient Com-
put. Intell., 8, 32-51.
36. Sarwar KM, Dey N & Ashour AS (2017). Large Scale Medical Data Mining
for Accurate Diagnosis: A Blueprint. In: Khan S, Zomaya A, Abbas A eds.
Handbook of Large-Scale Distributed Computing in Smart Healthcare. Sca-
lable Computing and Communications. Springer, Cham, pp. 157-176.
37. Singh A, Sharma A & Dey N (2015). Web recommendation Techniques-
Status, Issues and Challenges. Journal of Network Communications and
Emerging Technologies, 5, 57-65.
38. Santhi V (2017). Recent Advances in Applied Thermal Imaging for Indus-
trial Applications. Hershey: IGI Global.
39. Demongeot J, Hamie A, Glaria A & Taramasco C (2013). Dynalets : a new
time-frequency representation of periodic biological signals and spectral da-
ta. In IEEE AINA' 13. Piscataway: IEEE Press, pp. 1525-1532.
40. Demongeot J, Hamie A, Hansen O, Franco C, Sutton B & Cohen EP (2014).
Dynalets: a new method of modelling and compressing biological signals.
Applications to physiological and molecular signals. Comptes Rendus Biolo-
gies, 337, 609-624.
41. Demongeot J, Galli Carminati G, Carminati F & Rachdi M (2015). Stochas-
tic monotony signature and biomedical applications. Comptes Rendus Biolo-
gies, 338, 777-783.
42. Baconnier P, Benchetrit G, Demongeot J & Pham Dinh T (1984). Morpholo-
gie des Cycles Respiratoires : tentative de classification. In: Gallis JL, ed.,
Actes du IIIème Séminaire de Biologie Théorique, Bordeaux: Publications
University Bordeaux, pp. 87-102. (http://cat.inist.fr/?aModele=afficheN
&cpsidt =9132966).

41
43. Benchetrit G, Shea SA, Pham Dinh T, Bodocco S, Baconnier P & Guz A
(1989). Individuality of breathing patterns in adults assessed over time. Res-
piration Physiology, 75, 199-210.
44. http://home.ccr.cancer.gov/ncifdaproteomics/ppatterns.asp
45. Demongeot J, Hamie A, Laksaci A & Rachdi M (2016). Relative-Error Pre-
diction in Nonparametric Functional Statistics: Theory and Practice. Journal
of Multivariate Analysis, 146, 261-268.
46. Rachdi M, Laksaci A, Demongeot J, Abdali A & Madani F (2014). Theoreti-
cal and practical aspects of the quadratic error in the local linear estimation
of the conditional density for functional data. Computational Statistics & Da-
ta Analysis, 73, 53-68.
47. Preda C, Saporta G & Lévéder C (2007). Pls classification of functional data.
Computational Statistics, 22, 223-235.
48. http://www.who.int/whosis/icd10/ ; http://www.who.int/classifications/icf/en/
49. Demongeot J, Virone G, Duchêne F, Benchetrit G, Hervé T, Noury N & Ri-
alle V (2002). Multi-sensors acquisition, data fusion, knowledge mining and
alarm triggering in health smart homes for elderly people. Comptes Rendus
Biologies, 325, 673-682.
50. Virone G, Noury N & Demongeot J (2002). A system for automatic meas-
urement of circadian activity deviations in telemedicine. IEEE Trans. Bio-
med. Eng., 49, 1463-1469.
51. Rialle V, Duchêne F, Noury N, Bajolle L & Demongeot J (2002). Health
'smart' home: information technology for patients at home. Telemedicine
Journal and E-Health, 8, 395-409.
52. Le Bellego G, Noury N, Virone G, Mousseau M & Demongeot J (2006). A
Model for the Measurement of Patient Activity in a Hospital Suite. IEEE
Transactions ITB, 10, 92-99.
53. Vuillerme N, Chenu O, Pinsault N, Fleury A, Payan Y & Demongeot J
(2008). Can a plantar pressure-based tongue-placed electrotactile biofeed-
back improve postural control under altered vestibular and neck propriocep-
tive conditions? Neuroscience, 155, 291-296.
54. Vuillerme N, Pinsault N, Fleury A, Chenu O, Demongeot J, Payan Y & Pa-
van P (2008). Effectiveness of an electro-tactile vestibular substitution sys-
tem in improving upright postural control in unilateral vestibular-defective
patients. Gait Posture, 28, 711-715.
55. Vuillerme N, Pinsault N, Chenu O, Fleury A, Payan Y & Demongeot J
(2009). A wireless embedded tongue tactile biofeedback system for balance
control. Pervasive & Mobile Computing, 5, 268-275.
56. Chenu O, Payan Y, Hlavackova P, Demongeot J, Cannard F, Diot B &
Vuillerme N (2012). Pressure sores prevention for paraplegic people: effects
of visual, auditive and tactile supplementations on overpressures distribution
in seated posture. Applied Bionics and Biomechanics, 9, 61-67.
57. Fouquet Y, Franco C, Diot B, Demongeot J & Vuillerme N (2013). Estima-
tion of Task Persistence Parameter from Pervasive Medical Systems with
Censored Data. IEEE Trans. Mobile Computing, 12, 633-646.

42 J. Demongeot, M. Jelassi, C.Taramasco
58. Guttmann A, Ouchchane L, Li X, Perthus I, Gaudart J, Demongeot J & Boire
JY (2013). Performance map of a cluster detection test using the extended
power. International Journal of Health Geographics, 12, 47.
59. Franco C, Fleury A, Guméry P Y, Diot B, Demongeot J & Vuillerme N
(2013). iBalance-ABF: a Smartphone-Based Audio-Biofeedback Balance
System. IEEE Transactions on Biomedical Engineering, 60, 211-215.
60. Diot B, Hlavackova P, Demongeot J & Vuillerme N (2014). Sensory substi-
tution for balance control using a vestibular-to-tactile device. Multisensory
Research, 27, 313-336.
61. Rachdi M, Laksaci A, Demongeot J & Abdali A (2014). Theoretical and
practical aspects on the quadratic error in the local linear estimation of the
conditional density for functional data. Comp. Stat. Data Anal., 73, 53-68.
62. Demongeot J, Hansen O, Hamie A, Hazgui H, Virone G & Vuillerme N
(2015). Actimetry@home: actimetric tele-surveillance and tailored to the
signal data compression. In: ICOST’14, Lecture Notes in Comp. Sci., 8456,
59-70.
63. Demongeot J, Hamie A, Laksaci A & Rachdi M (2016). Relative-Error Pre-
diction in Nonparametric Functional Statistics: Theory and Practice. Journal
of Multivariate Analysis, 146, 261-268.
64. Demongeot J, Laksaci A, Naceri A & Rachdi M (2017). Local linear regres-
sion modelization when all variables are curves. Statistics and Probability
Letters, 121, 37-44.
65. Chan M, Campo E, Estève D & Fourniols JY (2009). Smart homes - current
features and future perspectives. Maturitas, 64, 90-97.
66. Demiris G & Hensel BK (2008). Technologies for an aging society: a sys-
tematic review of “smart home” applications. Yearb Med. Inform., 3, 33-40.
67. Shults MC, Rhodes RK, Updike SJ, Gilligan BJ & Reining WN (1994). A te-
lemetry instrumentation system for monitoring multiple subcutaneously im-
planted glucose sensors. IEEE Transactions on Biomedical Engineering,
41(10), 937-942.
68. Uchiyama A (1995). Endo-radioprobe needs micro machine technology. In
Proceedings of the 6th international symposium on micro machine and hu-
man science. Piscataway: IEEE Press, pp. 31-37.
69. Al-Taee MA, Al-Nuaimy W, Al-Ataby A, Muhsin ZJ & Abood SN (2015).
Mobile health platform for diabetes management based on the Internet-of-
Things. In Applied Electrical Engineering and Computing Technologies
(AEECT’15). Piscataway, IEEE Press, pp. 1-5.
70. Demongeot J, Noury N & Vuillerme N (2008). Data fusion for analysis of
persistence in pervasive actimetry of elderly people at home. In IEEE ARES-
CISIS' 08. Piscataway, IEEE Press, pp. 589-594.
71. Dittmar A, Axisa F, Delhomme G & Gehin C (2004). New concepts and
technologies in home care and ambulatory monitoring. Stud. Health Technol.
Inform., 108, 9-35.
72. http://www.legrand.fr/professionnels/les-appliques-ecoconfort-pour-
etablisse mentsmedicalises_2795.html

43
73. Demongeot J, Elena A, Jelassi M, Ben Miled S, Bellamine Ben Saoud N &
Taramasco C (2016). Smart homes: data fusion for preventive and therapy
education at home; example of type II diabetes. Information, 7, 50.
74. Couturier P, Franco A, Piquart JF, Mansotte J, Montani C, Suarez C, Mollier
A, Gucher C, Frossard M, Nicolas L, Jasso Mosqueda G, Mouchet MC, Ar-
gentier AM, Bosson JL, Carpentier P & Demongeot J (1996). "Telegerontol-
ogy": from teleassistance to teleconsultation of elderly people. Mythe or real-
ity ? Rev. Gériatrie, 21, 23-31.
75. Moreau-Gaudry A, Robineau F, André PF, Prince A, Pauget P, Demongeot J
& Payan Y (2006). Utilisation de la substitution sensorielle par électrostimu-
lation linguale pour la prévention des escarres chez les paraplégiques. Etude
préliminaire. L'escarre, 30, 24-37.
76. Vazquez-Buenos Aires J, Payan Y & Demongeot J (2003). Electrostimula-
tion of the tongue as a passive surgical guiding system. In de Almeida AT &
Nunes U (Eds.). ICAR’03. Piscataway: IEEE Press, 638-643.
77. Tassicker GE (1956). Preliminary report on a retinal stimulator. Br. J. Phys-
iol. Opt.,12, 102-105.
78. Starkiewicz W & Kuliszewski W (1963). The 80-channel elektroftalm. Pro-
ceedings of the lnternational Congress Technology Blindness. New York:
Am. Found. Blindness.
79. Konarzewski T (1992). 100-lecie elektroftalmu. Arch. Hist. Filoz. Med., 55,
95-99.
80. Visell Y (2009). Tactile sensory substitution: Models for enaction in HCI. In-
teracting with Computers, 21, 38–53.
81. Szkaradek M, Futyma-Ziaja M, Lewicka-Chomont A & Rejdak R 2014). A
Technology that Enables to Restore Visual Function. e-Polish J. Veterinary
Ophtalmol. Okulistyka, 3.
82. Fournier d'Albe EEF (1914). On a Type-Reading Optophone. Proceedings of
the Royal Society A: Mathematical, Physical and Engineering Sciences, 90
(619), 373–375.
83. Samsó Dies JM, Vila Bado J & Torres de Gasso JM (1962). Acoustic senso-
ry prosthesis for tactile stimulation. Animal experimentation. Plan and possi-
bilities of the same. Medicina Clínica, 39, 363-373.
84. Samsó Dies, Vila Bado J & Torres de Gasso JM (1965). Possibilités d’une
prothèse sensorielle acoustique par stimulation tactile Cybernetics of the
nervous system. In Wiener N & Schadé N (Eds.). Progress in Brain Research
17. Amsterdam: Elsevier, pp.169-178.
85. Bach-y-Rita P, Collins CC, Saunders FA, White B & Scadden L, (1969). Vi-
sion Substitution by Tactile Image Projection. Nature, 221, 963-964.
86. http://www.texisense.com/
87. http://www.feetme.fr/
88. Aud MA, Abbott CC, Tyrer HW, Neelgund RV, Shriniwar UG, Mohammed
A & Devarakonda KK (2010) Smart Carpet: Developing a Sensor System to
Detect Falls and Summon Assistance. J. Gerontological Nursing, 36, 8-12.

44 J. Demongeot, M. Jelassi, C.Taramasco
89. Tanaka O, Ryu T, Hayashida A, Moshnyaga VG & Hashimoto K (2014). A
Smart Carpet Design for Monitoring People with Dementia. Progress in Sys-
tems Engineering, 330, 653-659.
90. Andries M, Charpillet F & Simonin O (2015). High resolution pressure sens-
ing using sub-pixel shifts on low resolution load-sensing tiles. In ICRA’15.
Piscataway: IEEE Press, pp. 3890-3895.
91. http://www.bloomberg.com/bw/articles/2014-03-20/intellimat-flooring-
measures-health-based-on-footstep-patterns
92. Bourdessol H & Pin S (2008). Practice Guide Prevention of falls in the elder-
ly living at home. Paris: INPES.
93. Guettari T (2014). Détection de présence humaine et évaluation de la qualité
du sommeil en établissement d’hébergement pour personnes âgées dépen-
dantes (EHPAD). Paris: Télécom Sud Paris PhD Thesis.
94. Cavalcante Aguilar PA, Boudy J, Istrate D, Dorizzi B & Moura Mota IC
(2014). A Dynamic Evidential Network for Fall Detection. IEEE J. Biomedi-
cal & Health Informatics, 18, 1103-1113.
95. http://www.visuresp.com/
96. http://www.login.cl/noticias/1570-ssvq-y-universidad-de-valparaiso-
impulsan-innovador-proyecto-de-tele-rehabilitacion-para-pacientes-
cardiovasculares
97. Levitzky MG (2008). Using the pathophysiology of obstructive sleep apnea
to teach cardiopulmonary integration. Adv. Physiol. Educ., 32, 196-202.
98. Demongeot J & Demetrius L (2015). Complexity and Stability in Biological
Systems. Int. J. Bifurcation and Chaos, 25, 40013.
99. Demongeot J, Bezy-Wendling J, Mattes J, Haigron P, Glade N & J.L.
Coatrieux JL (2003). Multiscale Modeling and Imaging: The Challenges of
Biocomplexity. Proceedings of the IEEE Society, 91, 1723-1737.
100. Jelassi M (2017). Modélisation, simulation et analyse multiéchelle de ré-
seaux sociaux complexes : Application à l’aide à la prévention des maladies
contagieuses. PhD Thesis. Grenoble : University Grenoble Alpes.
101. Mall R, Rawi R, Ullah E, Kunji K, Khadir A, Tiss A, Abubaker J, Dehbi M
& Bensmail H (2017). Application of High-Dimensional Statistics and Net-
work based Visualization techniques on Arab Diabetes and Obesity data.
BioRxiv preprint, doi: http://dx.doi.org/10.1101/151621
102. Demongeot J, Jelassi M & Taramasco C (in Press). From Susceptibility to
Frailty in social networks: the case of obesity. Math. Pop. Studies, 24.