Context-Aware Personal Diet Suggestion System

brewerobstructionΤεχνίτη Νοημοσύνη και Ρομποτική

7 Νοε 2013 (πριν από 3 χρόνια και 9 μήνες)

64 εμφανίσεις

Context
-
Aware Personal Diet Suggestion System

Yu
-
Chiao

Huang
,

Ching
-
Hu Lu,
Tsung
-
Han Yang
,

and
Li
-
Chen Fu

Department of Computer Science and Information Engineering

National Taiwan University, Taipei, Taiwan


{
b9590
20
66,
D93922004,
D98922024
,
lichen}
@ntu.edu.tw

Abstract.

K
eeping a
health
y

and balanced
diet has
long been
a
critical
issue
for a person
wan
t-
ing
to stay fit and energetic
in her/his daily life
.
We can

always
turn to
a
dietitian
(or a
nutr
i-
tionist
)
for
professional
diet suggestion
s

if necessary
.

However,
we cannot have a dietitian
staying with us all
the time, which renders

daily nutrition

control

very
challenging
.

Therefore,
it
is
desirable

for each individual
to
receive
handy and informative

diet suggestion
s

whenever
necessary
. In this work, we propose
a
C
ontext
-
a
ware
P
ersonal Diet Suggestion System

(CPDSS)

which

tries to

maximize
an aggregated
health
utilit
y function

and
provide
s

useful diet sugge
s-
tion
s

according

to
s
ome

contextual
information
.
In order to increase the
practicality

of the pr
o-
posed system, we
have
integrate
d

the CPDSS with an everyday appliance

a smart
refriger
a-
tor

so that
we can
readily access

the
suggestions

about

one's

diet
,
receive

instant context
-
aware
reminders
while preparing foods
,
and

keep
long
-
term
diet history to
extract

more
useful
pa
t-
terns
to

caregivers
to
provide suggestion
s

for health
improve
ment
.

Keywords
:

context
-
aware
ness
, activity
level
, diet suggestion, smart
refrigerator

1.

Introduction

We

often

have the experience
of
sitting in front of
a

menu, unable to decide
a

meal to order
, let alone
making
a quick and wise

purchase

of

balanced
-
diet

foods
from

a

supermarket.
In addition
,
some factors including
personal preference
, food
intake

inhibition,
purchase
cost

of

food, advi
ce from professional medic care organ
i-
zation, are all important to our decision making. Apparently, a

smart

diet
-
suggestion
system that can help us figuring out these problems will be helpful, both
for

conve
n-
ience and
for
improving one

s health condition.

These suggestions are
especially

helpful for those who need
stricter

diet control such as children suffering from
Atte
n-
tion Deficit Hyperactivity Disorder

(ADHD)
, or elders suffering from high blood
pressure or other chronic
disease
s.

Most of
prior
researches

[1
-
5]

on diet suggest
ion
primarily
focus on proposing
some mechanisms, such as Linear Programming (LP), to achieve

nutrition or cost

optimization.

However, some other factors regarding human
-
related concerns are i
g-
nored, and most of these works perform well only when optimizin
g over cost or nutr
i-
tion limitation.
Reference

[6]

tries to
make use of
Case Base Reasoning to
implement
diet suggestion in hospital

to solve those human
-
involved problem
.


Due to the advances in sensing
technologies

and machine

learning,
we
propose

a
context
-
aware
Personal Diet Suggestion System

(abbreviated
C
PDSS
)

to take h
u-
man
-
related factors into consideration
.

T
he objective of the system, as its name su
g-






gests, is to help
a

user designing his or her individual diet menu

based on as many
available contexts as possible
.

T
he

contexts to help optimizing diet suggestion

includ
e

the
nutri
ent

of

a
food, the
diet
advisory

from

specialist
s, user
prior
preference
,
current

food
stocking
,
the
cost

of food,
the

diet
histor
ies of user
s
, and most importantly the
activity level inferred from ambient sensors.

By correctly inferring the personal acti
v-
ity level
, which is estimated based on the performed activities within a specific
inte
r-
val
;

we can dynamically sugges
t

menus
to
fulfill

nutri
tion needs

from
different

user
s

based on

their unique

concerns.

Due to the ever increasing
research

interest in
Smart Home and
Tele
-
h
ealthcare
,

m
ore and more
regular appliances are enhanced by the state
-
of
-
the
-
art ICTs and po
p-
ulated in our
daily

lives.

Some

of them
[7]

aim for
personalized

healthcare purposes.

One appliance is
the

so
-
called smart fridge
[
8
-
9]
, which

is

often
equipped with

bar
-
code scanners or RFID sensors

to h
elp
residents
easily monitor all stocking i
n-
formation like expi
ration date, nutrient contents
,

and
food
quantit
ies

via

a simple GUI
interface.

In order to ease information access

like most of prior works
,
this

work int
e-
grate
s

a smart

fri
d
ge as the main
user
interface
of the CPDSS
.
As an enhancement to
most of other
prior works
, our smart fridge is not a stand
-
alone appliance and it can
access and control other appliances via an integration platform in our smart
-
home
system.

2.

Context
-
Aware Personal Diet Suggestion System

2.1

System Overview

Fig.
1

illustrate
s

the proposed
CPDSS

where

an integration platform

[10]

can
integrate all
appliances, modules to

gather all context
s of interest
.

The
integration
platform
can provide information for remote access via the Internet (which
will be
implemented as an information cloud

using
Apache

Hadoop
)
.

Due
to
the
integration

platform,
our diet
suggestion

modules can be executed on a
RFID
-
equipped smart
fri
d
ge
and
in
terconnect with other
components

to access
various
contexts and provide
instant control
for the purpose of more satisfactory diet recommendation
. The sensors
deployed in our home environment can be used to
extract features for later
inference

of

ADL
s

(
acti
vities
of daily living)
[11]

and the inferred ADL
s

can be mapped to one's
activity level as an important context for diet suggestion.
The activity level performed
outdoors (such as jogging or walking, etc) ca
n be recorded/estimated by a mobile or a
wearable
device
, and the result stored in the device can be fed into the system once
the device can be accessed by the CPDSS
.

In order to illustrate the CPDSS, here
is

a brief scenario to

demonstrate the
benefits a user can obtain from
the
proposed
CPDSS.
Laura

is an octogenarian who
live
s alone

and her house has the proposed CPDSS
.
Since she

suffers from
some
ca
r-
diovascular

diseases such as high blood pressure
, she needs to pay more attention to
her diet
.

Each day when she wakes up, she click
s the
touch
screen on her smart fridge

to
get
a

suggested menu
for

her breakfast based on her
food

stocking.

Suggested by

a

doctor, she
need to ex
er
cise more than
one

hour
every morning

and do some chores
,
which
can be

correctly
detected

by
the
ADL
Inference Engine
.
One day, because
of
enough

exercise in the morning (her activity level is, say, high) and
the
hot
weather,






which

is a context provided by Other Context Generators,
t
he
CPDSS
therefore su
g-
gests

light
er

yet

higher
-
calor
ie food
in the menu at noon.
The
nutrition
-
aware
applic
a-
tion

can provide
a
reminder
via a speaker connected to the integration platform
to
inform

her
for

regular
replenish
ment of
water
or some
mineral
drink
. The system
can
check food stocking a
nd then
automatically

send
summarized
information

(including
a shopping list
)

to her families or a
healthcare center

for future reference
.

That afte
r-
noon,
Laura

s

activity level is low because of
her long

afternoon break, so the system
prepares a menu that
suggests

a menu with
fewer

calories and less nutrients reple
n-
ishment. The system also
automatically
avoids recommending high
-
sodium or
high
-
fat recipes according to the suggestion of
her
dietician.

In add
ition to

providing

diet
suggesti
ons
, the
CPDSS

can provide Laura's
diet
histories

via the

information
cloud
to share or exchange
useful information
if necessary
.


2.2

Problem Formulation

T
he main objective of
CPDSS
system is to improve
a

user

s
quality

of daily
living

via
balanced

diet
. For
easier evaluation

of
the system
, we define
one

utilit
y for
each context to be considered in our system
. Therefore, the objective
can be form
u-
lated as a problem

to
search
food

types

and their corresponding quantities to maxi
m-
ize
a

user

s
aggregate
d

utility
,
denoted as

U
, which is the
weighted

summation of
various individual utility functions
u
.

Before the formulation
,

we define some notation
s

first.

A food
f

is defined by a

nonnegative valued vector describing the nutrition it contains. That is,

f
= (

n
1
,

n
2
,

,
n
k

),

w
here
n
i

means the

i
th

nutrient
that

f
contains, assuming there are

k
nutrients at
total.

In addition,

n
i
(
f
) denote
s

the
i
th

nutrient of food
f

and

F

is

the
aggregate
d

set of
all foods.

With the definition of
F
, food type
T

is
a subset of
F,
which contains at least
one
kind of
food
.
Furthermore, the union of all

T
is exactly

F.

From the definition we
can infer that any
kind of
food in
F

must belongs to at least one food type.

For
si
m-
plicity,

we assume there are
h

food types.

We define

a
utility function
u

tak
ing

in a food
f

and
linearly
map
ping

it to a r
e-
al number

R
.

That is
,

u
(
f
1
) +
u
(
f
2
) =
u
(
f
1

+
f
2
)

and
u
(
q
f
) =
qu(
f
)
. With the individual
utility functions
,

w
e
calculate
an aggregate
d

utility
U
, which is simply the weighted
summation
of the functions,
and

all
weight
s

are

pre
-
configured.
U

is
the
objective

function that we wish to
optimize
.

To
be more specific
, assuming
we have a subset
F



Fig.
1
.

System Overview

CPDSS
@
Smart Home
ADL
Inference
Engine
Integration Platform
Other
Context
Generators
Nutrition
-
aware
Database
/
Applications
...
...
Sensors
/
Devices
Smart Phone

Smart
Refrig
.
Care
-
givers
The Internet
(
Cloud
)

Food
Source

Fitness
Center
Dieti
-
tian






of
F
, our objective
is

to decide a quantity for each food in
F

, which collectively

are

denoted
as

Q
.
That is, t
he solution will be consisted
of

F


and
Q
, a food list and their
corresponding
quantit
ies
.

The above
description

can be represented by:


1 2
1 2
,,...,
1
,,...,arg max ( ) where
,
m
m
m i i
q q q
i
Q q q q U q f F m


  
 
 
 


(
1
)

If we further
factorize

the
equation, we can get equation
s
(
2
)

and
(
3
)


1 2 1 2
,,...,,,...,
1 1
arg max ( ) arg max ( )
m m
m m
i i i i
q q q q q q
i i
Q U q f qU f
 

   

   
   
 

(
2
)


1 2
,,...,
1 1
arg max ( )
m
m p
i j j i
q q q
i j
Q q w u f
 

 
 

 
 
 
 
 

(
3
)

where
w

is a
weight
for

each utility function

and
there are

totally
p

utility functions.

To be more
succinct
, we define
a
matrix
Ω

so that
Ω
ij

=
u
i
(
f
j
), and
a matrix
W
,
which
contains
p

weights of all utility functions
. As a result
, the desired
Q*

is a
one

by
m

matrix which satisfies:






* arg max
T
Q
Q W Q
  

(
4
)

Note

that the matrix multiplication
W
Ω

can be pre
-
computed if
F


and
W

are

decided before the optimization. Therefore, the optimization problem is reduced to an
m

degree linear programmi
ng (LP) problem, where (
W
Ω
Q
T
) is exactly the objective
function that we want to optimize.

The c
onstraints in the
formation of a
LP problem
regarding diet suggestion
are
traditionally limited by
nutrition and food type
s
.
Unlike most of prior works,
the
CPDSS system first infers one's
activity level
or other contexts
so that the system
can
use
th
e
inf
ormation
as guidance
to

construct context
-
related

constraints
for

efficiently
searching the optimiz
ed solution

of
equation
(
4
)
.
In our work, each activity has
a pr
e-
defined score based on the characteristic of the activity. As a result, an activity level is
the
average

of summarized scores of all performed activities in a time interval
. Cu
r-
rently, we define
three
activity
levels

including LOW, MEDIUM, and HIGH.
W
e
pre
-
configure
th
e

statistics of daily allowance
[12]

of each nutrient and food type

given

each activity level

so that
w
e can
create

a lower bound and upper bound for
each nutrient and food type.

That is, different activity level
s

will
correspond to di
f-
ferent bounds
, leading to

different searching space.

For example, a user who exercises
a lot will definitely require more water and
nutrition
supply.

Mathematically speaking,
if
a

user

s activity level is
L
,

and

if
the

upper
-
bound
and
lower
-
bound of the

i
th

nutr
i-
ent allowance for
the

user
are denoted
by
n
i
+
(
L
)

and
n
i
-
(
L
)
,

we can express the nutrient
constraints by:


1 1
( ) ( ) and ( ) ( ), 1
m m
i j i j i j i j
i i
q n f n L q n f n L for j to k
 
 
  
 

(
5
)

Likewise
,

we can define
ρ
i
+
(L)

as the
upper
-
bound
of the
quantity of the
i
th

food type
under
a given

activity level
L
, and
ρ
i
-
(L)

is
the lower bound. Then we can have add
i-
tional

2
h

(
h

is the number of food types)
constraints

over food
quantit
ies
:


( ), and ( ), 1
i j i j
i j i j
f T f T
q L q L for j to k
 
 
 
  
 

(
6
)

For
generaliz
ation
, we can further add other
context
s

so that
different constraints
can
be
taken into consideration
.
Consequently
,

we append
a
parameter
C

to represent
the






additional information

to be considered
.

Note that
these extra

constraints can be
i
n-
crementally and selectively loosened or
dropped
according to their
significance

to
a
person of interest

if
no feasible
solution

can be

obtained
.

After
the
addition

of
these configurations, we have a complete LP problem
which has at most 2(
k
+
h
) constraints

to be optimized
.

2.3

System
Flowchart


As shown in
Fig.
2
, t
he
diet
suggestion

of the
CPDSS
can
be
primarily divided
into

sampling, solving, and combing
phase
s
.


A.

Sampling Phase

This phase

is responsible for selecting a
n

F


(a subset of foods) by
sampling a
food
-
searching space

to increase food variability
.
W
e first have an empty
F


and
then
sample
foods
in

stock

into
F

.
That is
, we randomly select food
s

to
F


so that there
are

at least some
predefined

portion of foods in each food type. That is, if there are 200
foods in
the
food type

starch


and the portion is 0.25
, there will be 50 foods selected
in
F

, including those foods in stock. The random selection can be
a
non
-
uniform di
s-
tribution. Th
e sampli
ng
can also
consider

other
information

including

user preference,
cost to get the food
s
, and personal
diet
-
inhibition.

B.

Solving Phase

When
the
previous phase outputs the selected
F’
, the remaining work will be
solving the LP problem
with

various

context
-
related constrains
.
Firstly, t
he activity
level (
L
)
, which is inferred by adding up
the

active degree
of
all
activit
ies

performed
by the user

in a specific interval
,

and other context
s of interest

(
C
) are integrated
to
determine

some upper/lower bounds of
the
LP
constraint
s
.
Secondly, we

select utility

function
s

for the
optimization

of the LP
. Even though the selection of
utility

functions

is flexible,
currently
we

primari
ly focus

on the following
three kinds

for
implication
,
they are
:




A
dvisory
utility
:

t
his
utility
is
based on

the evaluation of foods based on
advice

from nutrition specialist
s
,
and
is often the most relevant to
one

s health.
Intuitively,
t
hose foods that
may cause harmful
health
effect

will
get

negative utility values,
and positive values
if

beneficial

one
s

are chosen
.



Cost utility
:

this
utility includes the price of foods, the scarcity

of foods
, and
,

most
important
ly,

is whether the food is in
stock
.

This

utility

returns

a negative
value

since
it
is usually undesirable
to
get food
s at higher cost
,

so that its sign has to be
reversed before
calculation
.

A

food in stock will get positive cost utility (after sign
reversion) because
it is desirable to consume

refrigerator stocking
,

while other

foods

get

negative utility values

under the same
criterion
.



Preference utility
:

t
h
is

utility
is
user
-
dependent and may be

contradictory to
the
advisory utility. Patients who suffer obesity often prefer
high
-
calorie
and

high
-
fat
foods,
but

a
dietician may suggest the patient
s

stay

away from those
harmful
foods.
By default
, the weight
of

the
preference utility will always
be
lower than that of

Fig.
2
.

The
flowchart

of the CPDSS







the
advisory utility

in the CPDSS
.

For normalization,

th
e above

utility
values will not exceed
a predefined

range
.
For example
,
we can scale
all
of
our utility values
within

the interval
[
-
10, 10] after
normalization.

If
a user does not
like
the suggestion or
if
the
CPDSS
cannot
find

a

(primal)
sol
u-
tion

Q

from
F

,
the system

can
always re
-
initiate

the sampling phase

and

re
generate
another
F’
.
Since
there is randomization involved

in
the
sampling

phase
, it may
eventually

generate a feasible
F


so that
we

can
finally
find a solution
Q
. If it is still
not easy task to find a
Q
, we can try
gen
erating

a larger
F


wh
ile

in the
sampling
phase
so that there can be more candidate

foods to choose from
.

C.

C
ombining Phase

This phase

is an
on
-
going

yet

significant
task
for improving
system

practicality
.
I
n
the
previous two phase
s
,

we already have our solution: a food list
F


and its corr
e-
sponding quantity list
Q
,
but a user may feel confused if it comes to

combining
the
resultant

ingredients into
useful

and meaningful

recipes. More important,
a

recipe
itself also plays a role in
improving

one's

health

since

seasoning and
culinary

methods
are
also among critical issue of diet suggestion
, which will be discusse
d in
section 4

of the paper.


Fig.
3
.

Detailed
CPDSS
Fridge

Interface

(
can be more
simplified

for an elder
)







3.

System Implementation

We have implemented a prototype of
the
C
PDSS which
integrated
a
smart
fridge
as
its
user

interface.

The system
incorporates

the contexts regarding
a user's

activities
from
our
ADL
Inference Engine implemented by
Bayesian Networks

[11]
.
Based on the contexts, the system can in turn estimate

the activity level
of a user,
which represents the vigorousness of the user in an interval of interest
.
Currently,

the
CPDSS skip
s the
combining

phase

since its module is under development
.

In
this
prototype
,

we
have
equipped the fridge with an RFID reader
to
monitor

the stocking
status of
the fridge

as
shown in
Fig.
3
.

With

the touch panel of the smart fridge, w
e
can

display
summarized

histogram charts for users to visualize
their

related history
information
.


On the right
block
of the interface

lists
the inferred
activit
ies along with
the a
c-
tivity level

(namely, the
outputs

from the
ADL Inference Engine
) in a time interval
.
A

calendar

and

some
operation buttons are laid on the left of the interface.

The middle
block lists all
food in the fridge
. All the user has to do is to press down a button on the
left screen and the
diet suggestion

will be displayed in the bottom block

where

four
columns
indicat
e

their names, food types, and corresponding
quantit
ies
, and
the cu
r-
rent
stocking
status

as

a refe
rence for later
food
replenish
ment
. For simplicity, all
items that their corresponding quantities are below 0.25 units will be
ignored for o
p-
timization
.
After

pilot

simulation
s
, these items generally have corresponding
q

equal
10
-
9

units and are negligible
. In addition, the food quantity is approximated to mult
i-
ple of 0.5 for simplicity to users.

Currently,

the system

allow
s

a

user to set the weight
s

of three utility functions.
I
nitial setting of
the
weight
for
the
advisory utility
is

0.7, 0.2
for
the
cost
utility
,

and 0.1 for
the
preference utility.

In the
current
implementation,

foods
consist of

24 nutrients
including

calor
ie
,
protein
,
vitamin
,
and mineral
, etc
.

In addition, there
are

six

food types
, which are
starch
, protein (
e.g.

seafood, eggs, beans, and meats), dairy, oil (
e.g.

nuts and oil),
vegetable, and fruit.

In
this
implementation
,

the personal data includes user name,
suggested amount of
each nutrient from professionals


suggestion, and the lo
w-
er/upper bound
of nutrient
li
mitation
s
.

Table 1.

list
s

the simulated results suggested by the CPDSS
base
d

on a user's
Table 1. CPDSS Output Food Suggestion L
ist

with different Activity Level
s
.

Activity Level

LOW

HIGH

Food Suggestion

Wheat flakes, 2.5 units

Sweet potatoes, 0.5 units

Egg noodle, 1 unit

Cereal flake, 1.5 units

Bacon, 1 unit

Chicken liver, 2.5 units

Gallbladder liver
, 2.5 units

Low
-
fat milk powder, 2 units

Sunflower oil, 0.5 units

Low
-
calorie butter, 2.5 units

Agar, 2.5 units

Lavar, 2.5 units

Preserved longan, 2.5 units

Preserved starfruit, 2.5 units

Wheat
flakes, 1.5 units

Rice, 2.5 units

Sago Rice, 2.5 units

Buns, 0.5 units

Salmon floss, 2.5 units

Plumule bean noodle, 2 units

Salted eggs, 2.5 units

Full
-
fat milk powder, 2 units

Salad oil, 0.5 units

Low
-
calorie

butter, 2.5 units

Pumpkin, 2.5 units

Bamboo
shoots, 2.5 units

Orange, 2.5 units

Preserved pineapple, 2.5 units








LOW and HIGH
activity level
s for the last 24 hours
.
Apparently, the major differen
ce

is the total quantit
ies

of each food type.
In addition to
this
difference
, the suggestion
in fruit
type
is also
different
. T
he fruit with more water is weighted more when
one's
activity level changes from LOW to HIGH
(
in this case the orange

is suggested)
. The
suggestion
in the list of HIGH
activity

level
also
suggests
more
int
ake in
sodium
quantity

(in this case
2.5 units of
the
salted egg
s

are

suggested
)
.

Some may notice that the system chooses some scarce foods like liver or pr
e-
served foods. It may result from the wrong setting
s

of unit
-
weight
transformation

since

we
are still working on
the
unit
-
weight
transformation for the entire
food
dat
a-
base. As a result, those foods with small weight but high nutrition contents will be
more
likely

to be selected.

4.

Conclusion


In this paper we pr
opose a c
ontext
-
a
ware
p
ersonal
d
iet
s
uggestion
s
ystem

to
help one
improve

quality of
li
fe

through diet suggestion
. With the consideration of
various contexts (especially one's activity level), our goal is to provide healthier and
satisfactory diet suggestions. In
additional
, the CPDSS has
been integrated with an
integration platform so that the system has the potential for future enhancement to
cooperate with other appliances for better effect of a computer
-
aided diet regime.
Although it

is still a long way from
providing

diet suggestion
s

t
o actual health i
m-
provement,
the CPDSS is our first attempt toward more
promis
ing

balance
d

diet
.

The
functionality

of the system

can be further
extend
ed

in the future
if more information
can be gathered
or shared from outside of a smart home via a
cloud
-
co
mputing
based
services.

References

[1] Colavita, C. and D'orsi, R.: Linear programming and pediatric dietetics. British Journal of
Nutrition, 64, 02 (2007),pp. 307
-
317.

[2] Fletcher, L., Soden, P. and Zinober, A.: Linear programming tech
niques for the construction
of palatable human diets.
The Journal of the Operational Research Society
, 45, 5 (1994),pp.
489
-
496.

[3] Soden, P. and Fletcher, L.: Modifying diets to satisfy nutritional requirements using linear
programming.
British Journal o
f Nutrition
, 68, 03 (2007),pp. 565
-
572.

[4] Henson, S.: Linear programming analysis of constraints upon human diets.
Journal of
Agricultural Economics
, 42(1991),pp. 380
-
393.

[5] Chamberlin, J. A linear programming analysis of human diets based upon major s
taple crops:
sensitivity of low cost formulations to caloric, protein and bulk constraints. Massachusetts
Institute of Technology, Dept. of Mechanical Engineering, 1971.

[6] Khan, A. and Hoffmann, A.: Building a case
-
based diet recommendation system withou
t a
knowledge engineer.
Artificial Intelligence in Medicine
, 27, 2 (2003),pp. 155
-
179.

[7] Ryu, B. and Ryu, H.: Biological Measurement in Healthcare Refrigerator. City, (2009).

[8] Suhuai Luo, S., Jesse S. Jin, J. and Jiaming Li, J.: A Smart Fridge with an

Ability to
Enhance Health and Enable Better Nutrition.
International Journal of Multimedia and






Ubiquitous Engineering (IJMUE)
, 4, 2 (2009),pp. 69
-
79.

[9] Gu, H. and Wang, D.: A Content
-
aware Fridge Based on RFID in Smart Home for
Home
-
Healthcare. City, (2
009).

[10] Chun
-
Feng Liao, Y.
-
W. J., and Li
-
Chen Fu: Toward a Message
-
Oriented Application
Model and its Middleware Support in Ubiquitous Environments. In:
Proceedings of the 2008
International Conference on Multimedia and Ubiquitous Engineering
(Korea,200
8),pp.
298
-
303.

[11] Lu, C.
-
H. and Fu, L.
-
C.: Robust Location
-
Aware Activity Recognition Using Wireless
Sensor Network in an Attentive Home.
Automation Science and Engineering, IEEE
Transactions on
, 6, 4 (Oct. 2009),pp. 598
-
609.

[12] United
-
States
-
Recomme
nded
-
Daily
-
Allowances:
http://1stholistic.com/nutrition/hol_nutrition
-
RDA.htm
.