Ontology-based multi-agents for intelligent healthcare applications

ORIGINAL RESEARCH

Ontology-based multi-agents for intelligent healthcareapplications

Mei-Hui Wang • Chang-Shing Lee •

Kuang-Liang Hsieh • Chin-Yuan Hsu •

Giovanni Acampora • Chong-Ching Chang

Received: 14 July 2009 / Accepted: 20 February 2010 / Published online: 12 March 2010

� Springer-Verlag 2010

Abstract A healthy diet and lifestyle are the most

effective approaches to prevent disease. Good eating habits

are central to a healthy lifestyle. When a person eats too

much or too little on a continual basis, the risk of disease

will increase. Therefore, developing healthy and balanced

eating habits is essential to disease prevention. This paper

proposes an ontology-based multi-agents (OMAS), includ-

ing a personal knowledge agent, a fuzzy inference agent,

and a semantic generation agent, for evaluating the health

of diets. Using the proposed approach, domain experts can

create nutritional facts for common Taiwanese foods. Next,

the users are requested to input foods eaten. Finally, the

food ontology and personal profile ontology are con-

structed by domain experts. Fuzzy markup language (FML)

is used to describe the knowledge base and rule base of the

OMAS. Additionally, web ontology language (OWL) is

employed to describe the food ontology and personal

profile ontology. Finally, the OMAS semantically analyzes

dietary status for users based on the pre-constructed

ontology and fuzzy inference results. Using the generated

semantic analysis, people can obtain health information

about what they eat, which can lead to a healthy lifestyle

and healthy diet. Experimental results show that the pro-

posed approach works effectively and diet health status can

be provided as a reference to promote healthy living.

Keywords Ontology � Fuzzy inference � Agent �Fuzzy markup language (FML) � Diet planning

1 Introduction

Learning to know how to eat healthily is not easy for most

individuals. Therefore, an intelligent agent for planning

healthy diets is becoming an increasingly important

research topic. An intelligent agent that automatically

provides tips for how to choose foods that improve health

and avoid foods that increase risk of illness would be of

great assistance for most people, especially those with

diabetes or cardiovascular diseases. Wang et al. (2009)

proposed an intelligent healthy diet planning multi-agent

(IHDPMA) for healthy diet planning. Chen and Chen

(2008) noted that agent technology is a key area in the field

of artificial intelligence research and agents are being used

in an increasingly wide area of applications. For example,

Sanchez et al. (2009) presented the Semantic Web services

and Multi-Agent Systems framework (SEMMAS) for

seamless integration of technologies by using ontologies

that facilitate the interoperation of agents and Web ser-

vices. Zunino and Campo (2009) proposed a multi-agent

system called Chronos that assists users in organizing

meetings.

Each person has a unique diet. Therefore, to determine

individual dietary status, an ontology is a good idea for

building personal dietary patterns and providing nutritional

M.-H. Wang � C.-S. Lee (&)

Department of Computer Science and Information Engineering,

National University of Tainan, Tainan, Taiwan

e-mail: [email protected]

K.-L. Hsieh � C.-C. Chang

Graduate Institute of System Engineering,

National University of Tainan, Tainan, Taiwan

C.-Y. Hsu

Advance e-Commerce Institute,

Institute for Information Industry, Kaohsiung, Taiwan

G. Acampora

Department of Mathematics and Computer Science,

University of Salerno, Salerno, Italy

123

J Ambient Intell Human Comput (2010) 1:111–131

DOI 10.1007/s12652-010-0011-5

AJOU

Highlight

AJOU

Highlight

AJOU

Highlight

AJOU

Highlight

AJOU

Highlight

AJOU

Highlight

AJOU

Underline

AJOU

Highlight

AJOU

Underline

AJOU

Underline

facts for a healthy diet is needed. An ontology is a very

powerful tool for representing information and its seman-

tics; ontologies can be thought of as knowledge represen-

tations as ontologies do more than just control vocabulary

(Reformat and Ly 2009). Consequently, ontologies are

applied to many such research fields as news summariza-

tion (Lee et al. 2005) and medical information systems

(Orgun and Vu 2006). Furthermore, Reformat and Ly

(2009) proposed an ontology-based approach that provides

a rich environment for expressing different information

types, including perceptions. Multi-agent systems com-

bined with ontologies have been utilized to support dis-

tributed decision-making in such fields as manufacturing,

business, and engineering (Chen and Chen 2008). Lee et al.

(2006) presented a meeting scheduling system by combing

a genetic fuzzy agent with an ontology model. Debenham

and Sierra (2008) structured dialogues and process infor-

mation collected by agents using an agent communication

language.

Applying agent technology to healthcare is also an

important research topic because each description must

have a unique and clear healthcare meaning (Caceres et al.

2006). According to Moreno (2006), agent technology is

suitable for addressing healthcare problems for seven main

reasons. (1) An agent physically distributes required

knowledge to several locations. (2) An agent can model

autonomous entities. (3) Agent-based systems can employ

security mechanisms. (4) Agent technology has the ability

to communicate and coordinate. (5) Agents can automati-

cally discover and compose e-services. (6) Intelligent

agents have deliberate, reactive, and flexible behaviors and

can learn. (7) An agent’s autonomous, reactive, and flexible

characteristics make agents ideal for implementing ambient

intelligence applications. Many studies have applied agents

in healthcare, for example, respiratory waveform recogni-

tion (Lee and Wang 2007) and electrocardiogram appli-

cations (Lee and Wang 2008). Furthermore, the Multi-

agent Systems Group (GruSMA) team (Moreno et al. 2006)

designed and implemented a Healthcare Services multi-

agent system to help doctors reduce error at each diagnostic

and treatment stage. Caceres et al. (2006) developed a

semantic service-discovery mechanism that improves the

usability of a service-discovery system in medical emer-

gencies. Chan et al. (2008) created a multi-agent archi-

tecture for mobile health monitoring to check patient date,

reason collectively, and then recommend actions to

patients and medical staff in a mobile environment.

This paper combines ontology and agent technologies in

developing an intelligent multi-agent for diet evaluation. In

the proposed approach, domain experts first use Protege

software (Noy and McGuinness 2001) to pre-define a food

ontology according to nutritional facts obtained from the

Internet and convenient stores in Taiwan. Additionally,

ontology domain experts pre-built the personal profile

ontology for users. The proposed ontology-based multi-

agents (OMAS), including a personal knowledge agent, a

fuzzy inference agent, and a semantic generation agent,

provides users with a semantic health-based description of

the foods they eat. A fuzzy markup language (FML) is

utilized to model the knowledge base and rule base of the

OMAS (Acampora and Loia 2005a, b). Finally, semantic

results for healthy diet status are stored in the healthy diet

status repository to provide users with a reference for

future dietary choices. The remainder of this paper is

organized as follows: Section 2 describes the ontology

model for semantic healthcare applications. Section 3

presents the applications of fuzzy markup language (FML)

for healthcare. Section 4 introduces the ontology-based

multi-agents for healthy diet evaluation. Section 5 gives

some experimental results. Section 6 provides conclusions.

2 Ontology model for semantic healthcare applications

This section describes the ontology model for semantic

healthcare applications. Section 2.1 briefly introduces the

structure of the semantic healthcare application. Sec-

tions 2.2 and 2.3 describe the domain ontology model and

personal profile ontology, respectively.

2.1 Structure of the semantic healthcare application

Figure 1 shows the structure of the semantic healthcare

applications. A three-layer structure with a knowledge

layer, communication layer, and application layer, is

employed for the semantic healthcare applications.

The knowledge layer has a knowledge base, a rule base,

a personal profile ontology, a food ontology, and a healthy

diet status repository. The communication layer provides

application interfaces, such as an FML and web ontology

language (OWL), for interacting with the application layer

and knowledge layer. The FML describes the knowledge

base using fuzzy concepts and fuzzy rules of the OMAS;

the OWL describes the food ontology and personal profile

ontology. First, a user with a personal digital assistant

(PDA), personal computer, or notebook computer accesses

the OMAS platform via the Internet. After authentication, a

user can access the OMAS platform. Next, a user inputs the

foods eaten. Finally, based on knowledge stored in the

knowledge layer, a user can receive an analysis of dietary

status via the communication layer.

2.2 Domain ontology model

Based on the levels of organization (Ferber 1999) and

previous work (Lee et al. 2005), this paper presents a

112 M.-H. Wang et al.

123

AJOU

Underline

AJOU

Underline

AJOU

Underline

AJOU

Highlight

AJOU

Highlight

AJOU

Typewriter

Agent technology가 헬스케어 도메인의 문제를 해결할 수 있는 7가지 이유

AJOU

Highlight

AJOU

Underline

AJOU

Rectangle

domain ontology model for healthcare applications. Fig-

ure 2 shows the structure of the domain ontology model,

which has a domain name, category set, and concept set.

The domain name represents the name of the ontology

model. The category set contains several categories,

labeled ‘‘category 1, category 2, category 3, …, and cate-

gory k.’’ Each concept in the concept set contains a concept

name Cm and an attribute set ACm1; � � � ;ACmqm

� �for an

application domain. In addition, a relationship exists

among concepts that belong to the same category. For

Food OntologyPersonal Profile

Ontology

Internet

Personal MealInput

OMAS Platform

Application Layer

Communication Layer

Knowledge Layer

KnowledgeBase

Rule Base

Healthy Diet State Repository

Note:

(1) OMAS: Ontology-based Multi-AgentS (2) FML: Fuzzy Markup Language (3) OWL: Web Ontology Language

OntologyDomain Experts

OMAS

FML OWL

HealthcareDomain Experts

Fig. 1 Structure of semantic healthcare applications

Concept Set

Category Set

C1

C3

Cm

Domain Name

… …

…

…

……

…

Category 1

Category k

Category 3

Category 2

Acm2Acm1

Acmqm

Acm3

Ac12

Ac1q1

Ac11

C 2

…

Ac22Ac23

Ac2q2

Ac21

Ac31

Ac32Ac33

Ac3q3

C4

…

Ac42Ac 43

Ac4q4

Ac41

C5

…

Ac52Ac53

Ac5q5

Ac 51

Ac13

Fig. 2 Structure of the domain ontology

Ontology-based multi-agents for intelligent healthcare applications 113

123

example, there is a bi-directional arrow between concepts

C1 and C2 because these two concepts belong to the

category 2. We assume a meal comprises a main course,

side dish, dessert, and beverage. Moreover, each compo-

nent of each meal has its own nutritional facts, which can

contain product-specific information such as serving size,

calories, and nutritional information based on the nutri-

tional facts on food labels (US Food and Drug Admin-

istration 2009).

Figure 3 shows the constructed food ontology. The

domain name of the food ontology is Meal. Categories in

the category set are Main Courses, Side Dishes, Desserts,

and Beverages. The concept set contains many concepts. In

this paper, calories per portion as well as grams of carbo-

hydrates, protein, and fat per portion are considered in the

constructed food ontology. Therefore, each concept has

four attributes—Calories, Carbohydrates, Protein, and

Fat. Additionally, each attribute has an attribute value. The

food ontology indicates that (1) Main Courses, Sides

Dishes, Desserts, and Beverages, are part of a meal.

(2) Concepts Citizen Lunch Box and Hawaiian Pizza are

under the Main Courses category. (3) Concepts Coca-Cola

and Apple Juice belong to the Beverages category. (4)

Concepts Egg Pudding and Coffee Egg Roll are Desserts

category. (5) Concepts Garlic Cheese Bread and Grapes

Toast are under Side Dishes category. (4) The Carbohydrate,

Protein, Fat, and Calories are 93, 26, 29 g, and 737 kcal,

respectively, for the concept Citizen Lunch Box. (5) The

values of attributes Carbohydrate, Protein, Fat, and Cal-

ories are 74, 13, 11 g, and 445 kcal, respectively, for the

concept Grapes Toast. (6) The attribute values of Carbo-

hydrate, Protein, Fat, and Calories are 17, 6, 16 g, and

236 kcal, respectively, for Egg Pudding. (7) Concepts,

belonging to the same category, exist a relationship, for

example, both concepts Coca Cola and Apple Juice are the

Beverages category; therefore, a bi-directional arrow is

shown between these two concepts.

2.3 Personal profile ontology

Individuals vary in profiles and eating habits. A personal

profile can be based on factors such as an individual’s

gender, age, height, and weight. Thus, personal daily

planned calories must be based on body mass index (BMI).

Notably, as a meal can comprise a main course, side dishes,

desserts, and beverages, people typically only know what

they eat, not whether they eat healthily. Additionally,

according to nutritional information, each dish has grams

of carbohydrates, protein, and fat per portion, such that

Actual Calories a dish contains per portion can be known.

With the assistance of nutritional information, the Per-

centage of Calories from Carbohydrate (PCC), Percentage

Meal

Citizen LunchBox

Hawaii Pizza

Carbo-hydrate

93 g

Fat29 g

Calories737 kcal

Protein10 g Fat

12 g

Calories336 kcal

Carbo-hydrate

47 g

Garlic CheeseBread

Grapes Toast

MainCourses Dess-

erts

Bever-ages

Carbo-hydrate

17 g

Protein0 g

Protein6 g

Carbo-hydrate

66 g

Fat16 g

Fat0 gCalories

236 kcal

Calories252 kcal

Egg Pudding

CoffeeEgg Roll

Coca Cola

Apple Juice

Protein26 g

Category SetSide

Dishes

Concept Set

Fat13 g

Calories265 kcal

Carbo-hydrate

31 gProtein17g

Fat11 g

Calories445 kcal

Carbo-hydrate

74 gProtein13 g

Carbo-hydrate

17 gProtein

2 gFat10 g

Calories166 kcal

Carbo-hydrate

12 gProtein

0 gFat0 g

Calories46 kcal

...

...

...

...

...

...

...

...

...

...

...

Fig. 3 Structure of the food ontology


123

of Calories from Protein (PCP), Percentage of Calories

from Fat (PCF), and Caloric Difference (CD) between

actual calories people consume and planned caloric intake

set by nutritionists, are acquired. Based on the above

mentioned information, the personal profile ontology is

constructed. Figure 4 shows an example of the user’s

personal profile ontology. This person’s sex is male, Age is

30, Height is 170 cm, Weight is 60 kg, BMI is 20.7 kg/m2,

PCP is 8%, PCC is 57%, PCF is 36%, and CD is 156 kcal.

3 Fuzzy Markup Language for healthcare applications

Fuzzy Markup Language (FML) is a fuzzy-based markup

language that can manage fuzzy concepts, fuzzy rules, and

a fuzzy inference engine (Acampora and Loia 2005a, b).

Additionally, FML is composed of three layers—eXtensi-

ble Markup Language (XML), a document-type definition,

and extensible stylesheet language transformations. Sec-

tion 3.1 gives an FML overview. Section 3.2 introduces an

FML implementation, and Section 3.3 introduces the FML-

Based intelligent healthcare application. Some examples of

the OMAS are shown in Section 3.4.

3.1 Fuzzy Markup Language overview

Differently from other similar approaches such as Fuzzy

Control Language (FCL) or MATLAB Fuzzy Inference

System (FIS) developed by the Matworks, FML exhibits

additional benefits in the FLC programming that are related

to its XML derivation. Indeed, whereas FCL or FIS code is

totally based on a textual representation, FML programs

are coded through a collection of correlated semantic tags

capable of modeling different controller components by

exploiting abstraction benefits offered by XML tools. From

the implementation point of view, these benefits allow

fuzzy designers (1) to simply code their ideas on hetero-

geneous hardware without having knowledge about pro-

gramming details related to the different platforms and (2)

to program a fuzzy controller without referring to general

purpose computer languages. In our scenario, FML allows

us to program diet agents in a very short time and simul-

taneously to test the inferred results on different hardware

without additional programming works.

In this sub-section, FML is explained starting from a

very high level introduction to fuzzy systems, the most

wide-spread application of fuzzy logic. Indeed, fuzzy logic

controllers (FLCs) have been applied in consumer goods

such as video cameras and washing machines, or industrial

processes such as cement kilns and trains. Since Zadeh’s

coining of the term fuzzy logic and Mamdani’s early

demonstration of FLC, enormous progress has been made

by the scientific community in the theoretical, as well as

application fields of FLC. Trivially, a fuzzy control allows

the designer to specify the control in terms of sentences

rather than equation by replacing a conventional controller,

Subject 1

Meal No. 1

PersonalMeal

PersonalInformation

Age

Category Set

Concept Set

Value30

...

...

Height

Value170cm

Weight

Value60kg

BMI

Value20.7kg/m^2

Planned Calories_1

Value584kcal

Sex

ValueMale

Value156kcal

CD_1

Carbo-hydrate

105 g

Fat30 g

Calories740 kcalProtein

14 g

Value8%

Value57%

Value36%

PCP_1 PCC_1 PCF_1

Meal No. n

Carbo-hydrate

119 g

Fat29 g

Calories857 kcalProtein

30 g

BMI: Body Mass IndexPCC: Percentage of Calories from CarbohydratePCP: Percentage of Calories from ProteinPCF: Percentage of Calories from FatCD: Caloric Difference

...... ...

... ...

...

......

... ...

...............

Fig. 4 Structure of the personal profile ontology


123

a proportional-integral-derivative (PID) controller, with

linguistic IF–THEN rules. The use of linguistic variables

represents a significant paradigm shift in system analysis:

using the linguistic approach, the focus of the attention in

the dependencies representations is shifted from differen-

tial equations to fuzzy IF–THEN rules in the form if X is A

then Y is B. The propositions X is A and Y is B are called

fuzzy clauses; A and B are linguistic variables and X and Y

are their linguistic values (for instance, if Pressure is high

then Volume is low). The main components of fuzzy con-

troller are (organized in a hierarchical way): (1) fuzzy

knowledge base; (2) fuzzy rule base; (3) inference engine;

(4) fuzzification subsystem; and (5) defuzzification

subsystem.

3.2 Fuzzy Markup Language Implementation

An FML implementation is based on the employment of

tags able to model different parts of fuzzy controller in a

taxonomic way. The root of fuzzy controller taxonomy, the

Controller node, is represented through the FML tag

\FUZZYCONTROL[. Such tag represents the root tag of

FML programs, that is, the opening tag of each FML

program. \FUZZYCONTROL[ uses three tags: type,

defuzzyfyMethod, and ip. The type attribute permits to

specify the kind of fuzzy controller, Mamdani or Takagi–

Sugeno–Kang (TSK); defuzzyfyMethod attribute defines the

defuzzification method used in modeled controller; ip is

used to define the location of controller in the computer

network. The fuzzy knowledge base is defined by means of

the tag \KNOWLEDGEBASE[ which maintains the set

of fuzzy concepts used to model the fuzzy rule base. The

\KNOWLEDGEBASE[ uses the attribute ip that deter-

mines the location in the network of whole fuzzy knowl-

edge base of our system. In order to define the fuzzy

concept related controlled system, \KNOWLEDGE-

BASE[ tag uses a set of nested tags: (1) \FUZZYVARI-

ABLE[; (2)\FUZZYTERM[; (3) a set of tags defining a

shape of fuzzy sets.

\FUZZYVARIABLE[ defines the fuzzy concept, for

example, ‘‘temperature’’; \FUZZYTERM[ defines a lin-

guistic term describing the fuzzy concept, for example,

‘‘low temperature’’; the set of tags defining the shapes of

fuzzy sets are related to fuzzy terms. The attributes of

\FUZZYVARIABLE[ tags are: name, scale, domainLeft,

domainRight, type, and ip. (1) The name attribute defines

the name of fuzzy concept, for instance, temperature; (2)

scale is used to define the scale used to measure the fuzzy

Fig. 5 Knowledge base and rule base of the intelligent healthcare applications


123

AJOU

Highlight

AJOU

Highlight

AJOU

Highlight

AJOU

Highlight

AJOU

Highlight

AJOU

Highlight

AJOU

Highlight

concept, for instance, Celsius degree; (3) domainLeft and

domainRight are used to model the universe of discourse of

fuzzy concept, that is, the set of real values related to fuzzy

concept; (4) the position of fuzzy concept into rule (con-

sequent part or antecedent part) is defined by type attribute;

(5) ip defines the position of fuzzy knowledge base in the

Table 1 Part of the FML view of the OMAS

<?xml version="1.0" encoding="UTF-8" standalone="no"?> <FuzzyController ip="140.133.33.2" name="Healthy Diet"> <KnowledgeBase> <FuzzyVariable domainleft="0.0" domainright="100.0" name="PCC" scale="%" type="input"> <FuzzyTerm complement="false" name="Low"> <TrapezoidShape Param1="0.0" Param2="0.0" Param3="50.0" Param4="55.0"/> </FuzzyTerm> <FuzzyTerm complement="false" name="Balanced"> <TrapezoidShape Param1="50.0" Param2="55.0" Param3="65.0" Param4="70.0"/> </FuzzyTerm> <FuzzyTerm complement="false" name="High"> <TrapezoidShape Param1="65.0" Param2="70.0" Param3="100.0" Param4="100.0"/> </FuzzyTerm> </FuzzyVariable> <FuzzyVariable domainleft="0.0" domainright="250.0" name="CD" scale="%" type="input"> <FuzzyTerm complement="false" name="Acceptable"> <TrapezoidShape Param1="0.0" Param2="0.0" Param3="50.0" Param4="100.0"/> </FuzzyTerm> <FuzzyTerm complement="false" name="MoreOrLessAcceptable"> <TrapezoidShape Param1="70.0" Param2="100.0" Param3="150.0" Param4="200.0"/> </FuzzyTerm> <FuzzyTerm complement="false" name="UnAcceptable"> <TrapezoidShape Param1="150.0" Param2="200.0" Param3="250.0" Param4="250.0"/> </FuzzyTerm> </FuzzyVariable> ... <FuzzyVariable accumulation="MAX" defaultValue ="0.0" defuzzifier="COG" domainleft="0.0" domainright="1.0" name="HDS" scale="" type="output"> <FuzzyTerm complement="false" name="VeryUnHealty"> <TrapezoidShape Param1="0.0" Param2="0.0" Param3="0.1" Param4="0.25"/> </FuzzyTerm> <FuzzyTerm complement="false" name="UnHealthy"> <TrapezoidShape Param1="0.1" Param2="0.25" Param3="0.25" Param4="0.5"/> </FuzzyTerm> <FuzzyTerm complement="false" name="MediumHealthy"> <TrapezoidShape Param1="0.25" Param2="0.5" Param3="0.5" Param4="0.75"/> </FuzzyTerm> <FuzzyTerm complement="false" name="Healthy"> <TrapezoidShape Param1="0.5" Param2="0.75" Param3="0.75" Param4="0.9"/> </FuzzyTerm> <FuzzyTerm complement="false" name="VeryHealthy"> <TrapezoidShape Param1="0.75" Param2="0.9" Param3="1.0" Param4="1.0"/> </FuzzyTerm> </FuzzyVariable> </KnowledgeBase> <RuleBase activationMethod="MIN" andMethod="MIN" name="RuleBase1" orMethod="MAX" type="mamdani"> <Rule connector="and" name="RULE1" operator="MIN" weight="1.0"> <Antecedent> <Clause> <Variable>PCC</Variable> <Term>Low</Term> </Clause> <Clause> <Clause> <Variable> CD</Variable> <Term>Acceptable</Term> </Clause> </Antecedent> … <Consequent> <Clause> <Variable> HDS</Variable> <Term>UnHealthy</Term> </Clause> </Consequent> … </RuleBase> </FuzzyController>


123

AJOU

Highlight

AJOU

Highlight

AJOU

Highlight

AJOU

Highlight

AJOU

Highlight

Table 2 Part of the constructed fuzzy rules of the OMAS

Rule no. Input fuzzy variables Output fuzzy variable

Age BMI PCC PCP PCF CD HDS

R1 Young UnderWeight Low Low Low Acceptable UnHealthy

R2 Young UnderWeight Low Low Low M.L.UnAcceptable UnHealthy

R3 Young UnderWeight Low Low Low UnAcceptable VeryUnHealthy

R4 Young UnderWeight Low Low Balanced Acceptable UnHealthy

R5 Young UnderWeight Low Low Balanced M.L.UnAcceptable UnHealthy

R6 Young UnderWeight Low Low Balanced UnAcceptable VeryUnHealthy

R7 Young UnderWeight Low Low High Acceptable UnHealthy

R8 Young UnderWeight Low Low High M.L.UnAcceptable UnHealthy

R9 Young UnderWeight Low Low High UnAcceptable VeryUnHealthy

R10 Young UnderWeight Low Balanced Low Acceptable UnHealthy

R11 Young UnderWeight Low Balanced Low M.L.UnAcceptable UnHealthy

R12 Young UnderWeight Low Balanced Low UnAcceptable VeryUnHealthy

R13 Young UnderWeight Low Balanced Balanced Acceptable Healthy

R14 Young UnderWeight Low Balanced Balanced M.L.UnAcceptable MediumHealthy

R15 Young UnderWeight Low Balanced Balanced UnAcceptable VeryUnHealthy

R16 Young UnderWeight Low Balanced High Acceptable UnHealthy

R17 Young UnderWeight Low Balanced High M.L.UnAcceptable UnHealthy

R18 Young UnderWeight Low Balanced High UnAcceptable VeryUnHealthy

R19 Young UnderWeight Low High Low Acceptable UnHealthy

R20 Young UnderWeight Low High Low M.L.UnAcceptable UnHealthy

… … … … … …R701 Old OverWeight Balanced High High M.L.UnAcceptable UnHealthy

R702 Old OverWeight Balanced High High UnAcceptable VeryUnHealthy

R703 Old OverWeight High Low Low Acceptable UnHealthy

R704 Old OverWeight High Low Low M.L.UnAcceptable UnHealthy

R705 Old OverWeight High Low Low UnAcceptable VeryUnHealthy

R706 Old OverWeight High Low Balanced Acceptable MediumHealthy

R707 Old OverWeight High Low Balanced M.L.UnAcceptable UnHealthy

R708 Old OverWeight High Low Balanced UnAcceptable VeryUnHealthy

R709 Old OverWeight High Low High Acceptable UnHealthy

R710 Old OverWeight High Low High M.L.UnAcceptable UnHealthy

R711 Old OverWeight High Low High UnAcceptable VeryUnHealthy

R712 Old OverWeight High Balanced Low Acceptable MediumHealthy

R713 Old OverWeight High Balanced Low M.L.UnAcceptable UnHealthy

R714 Old OverWeight High Balanced Low UnAcceptable VeryUnHealthy

R715 Old OverWeight High Balanced Balanced Acceptable Healthy

R716 Old OverWeight High Balanced Balanced M.L.UnAcceptable MediumHealthy

R717 Old OverWeight High Balanced Balanced UnAcceptable VeryUnHealthy

R718 Old OverWeight High Balanced High Acceptable UnHealthy

R719 Old OverWeight High Balanced High M.L.UnAcceptable UnHealthy

R720 Old OverWeight High Balanced High UnAcceptable VeryUnHealthy

R721 Old OverWeight High High Low Acceptable UnHealthy

R722 Old OverWeight High High Low M.L.UnAcceptable UnHealthy

R723 Old OverWeight High High Low UnAcceptable VeryUnHealthy

R724 Old OverWeight High High Balanced Acceptable UnHealthy

R725 Old OverWeight High High Balanced M.L.UnAcceptable UnHealthy

R726 Old OverWeight High High Balanced UnAcceptable VeryUnHealthy


123

Fig. 6 Surface view for OMAS a Inputs (X-PCP and Y-PCC) output

(HDS), b Inputs (X-PCF and Y-PCP) output (HDS), c Inputs (X-PCFand Y-PCC) output (HDS), d Inputs (X-CD and Y-PCC) output

(HDS), e Inputs (X-CD and Y-PCP) output (HDS), and f Inputs

(X-CD and Y-PCF) output (HDS)

Table 2 continued

Rule no. Input fuzzy variables Output fuzzy variable

Age BMI PCC PCP PCF CD HDS

R727 Old OverWeight High High High Acceptable UnHealthy

R728 Old OverWeight High High High M.L.UnAcceptable UnHealthy

R729 Old OverWeight High High High UnAcceptable VeryUnHealthy


123

computer network. \FUZZYTERM[ uses one attribute,

name used to define the linguistic value associate with

fuzzy concept. Fuzzy shape tags, used to complete the

definition of fuzzy concept, are: \TRIANGULAR-

SHAPE[, \LINEARSHAPE[, \TRAPEZOIDSHAPE[,

\SSHAPE[, \ZSHAPE[, and \PISHAPE[. Every

shaping tag uses a set of attributes which defines the real

outline of corresponding fuzzy set. The number of these

attributes depends on the chosen fuzzy set shape.

The root of fuzzy rulebase component is modeled by the

\RULEBASE[ tag which defines a fuzzy rule set. The

\RULEBASE[ tag uses two attributes: inferenceEngine

and ip. The former is used to define inference operator

type: MinMaxMinMamdani or LarsonProduct. The latter

defines the network location of the set of rules used in

fuzzy controller. In order to define the single rule, the

\RULE[ tag is used. The tags used by \RULE[ are: id,

connector, weight, and ip. (1) The id attribute permits to

identify the rule; (2) connector is used to define the logical

operator used to connect the different clauses in antecedent

part; (3) weight defines the importance of rule during

inference engine time; (4) ip defines the location of rule in

the computer network. The definition of antecedent and

consequent rule part is obtained by using \ANTECED-

ENT[ and \CONSEQUENT[ tags. \CLAUSE[ tag is

used to model the fuzzy clauses in antecedent and conse-

quent part. In order to treat the fuzzy operator ‘‘not’’ in

clauses, the tag \CLAUSE[ uses the boolean attribute

‘‘not’’. To complete the definition of fuzzy clause, the

\VARIABLE[, \TERM[, and \TSKPARAM[ have to

be used. In particular, the pair ‘‘\VARIABLE[,

\TERM[’’ is used to define fuzzy clauses in antecedent

and consequent part of Mamdani controllers rules as well

as in antecedent part of TSK controllers rules. While, the

pair ‘‘\VARIABLE[, \TSKPARAM[’’ is used to model

the consequent part of TSK controllers rules.

3.3 FML-based intelligent healthcare application

Based on the FML, an FML editor, developed by the LASA

Laboratory, University of Salerno, Italy, is used to construct

the knowledge base and rule base of the OMAS (Lee et al.

2009; Wang et al. 2009). The knowledge base describes

fuzzy concepts of the OMAS, including fuzzy variables,

fuzzy terms, and membership functions of fuzzy sets. Con-

versely, the rule base describes the fuzzy rule set, including

the antecedent and consequent rules. Figure 5 shows the

knowledge base and rule base of the OMAS in FML.

Table 1 lists the part of FML view of intelligent

healthcare applications (variables and rules definition). It

has one output fuzzy variable Healthy Diet Status (HDS),

729 fuzzy rules, and six input fuzzy variables—Age, BMI,

PCC, PCP, PCF, and CD. Each fuzzy variable has several

fuzzy terms. For example, fuzzy variable Age has three

ConvenientStore

OntologyDomain Experts

FML Editor

Protege

FML

OWLOntology

Repository

Personal Meal Input

Internet

Ontology-based Multi-AgentS (OMAS)

Healthy Diet Status Repository

Food

Personal Knowledge

Agent

FuzzyInference

Agent

SemanticGeneration

Agent

HealthcareDomain Experts

Utilization Phase

Construction Phase

Fig. 7 Structure of the OMAS


123

fuzzy terms, namely, Young, Middle, and Old. In particular,

our systems uses 729 fuzzy rules corresponding to all

possible combinations of input terms (#terms#vari-

ables = 36 = 729). The ideas behind the construction of the

fuzzy rules are as follows. From the view point of the

nutritionists, the food eaten is healthy or not depends on the

amount of the carbohydrate, protein, and fat, as well as

difference between actual calories people consume and

planned caloric intake set by nutritionists. Additionally,

based on the personal height, weight, age, and individual

lifestyle, the nutritionists set the planned caloric intake for

each person and the BMI is acquired. Therefore, Age, BMI,

PCC, PCP, PCF, and CD are considered as the input fuzzy

variables of the OMAS FML, and then HDS is obtained

after the fuzzy inference mechanism. The healthier the

food eaten is (1) the more balanced the value PCC, PCF,

and PCF are; as well as (2) the closer to the planned caloric

intake the actual calories people consume. Hence, the

ontology domain experts constructed the fuzzy rules

according to these criteria and Table 2 lists part of the

constructed fuzzy rules of the OMAS, where M. L. UnAc-

ceptable means MoreOrLessAcceptable.

3.4 Example of the OMAS

Based on the constructed fuzzy rules built by the ontology

domain experts, the output 3-dimensional surfaces are

constructed based on the OMAS. Figure 6 represents the

surface view for OMAS.

Major trends can be observed in the following rule

descriptions. (1) The HDS does increase as the PCC

approaches the balanced state (roughly 55–65%). (2) The

HDS does increase as the PCP approaches the balanced

state (roughly 10–20%). (3) The HDS does increase as PCF

approaches the balanced state (about 25–35%). (4) The

HDS does increase as the CD approaches the acceptable

state (about 0–50 kcal). (5) If PCC is balanced, PCP is

balanced, PCF is balanced, and CD is acceptable, then

HDS increases. Based on the major trends (Fig. 6), the

behavior of the proposed OMAS meets the nutritional

principles; that is, eating balanced foods improves one’s

health. Hence, the proposed approach is feasible for heal-

thy diet evaluation, which will help people create healthy

eating habits.

4 Ontology-based multi-agents for healthy diet

evaluation

This section introduces the ontology-based multi-agents

(OMAS) for healthy diet evaluation. Section 4.1 briefly

describes of the OMAS structure, which has three agents—

a personal knowledge agent, a fuzzy inference agent, and a

semantic generation agent. The details of each agent are

given in Sections 4.2, 4.3, and 4.4.

4.1 Structure of the ontology-based multi-agents

Figure 7 shows OMAS structure, including a construction

phase and a utilization phase, and the OMAS operates as

follows:

• Construction Phase: (1) Nutritional information of

foods is gathered from the Internet and convenient

stores in Taiwan. (2) There are two kinds of domain

experts, including ontology domain experts and health-

care domain experts, involving the experiments shown

in this paper. The ontology domain experts are

responsible for building ontology and setting up rules.

The healthcare domain experts are to evaluate the

performance of the proposed approach. (3) The ontol-

ogy domain experts then use Protege software (Noy and

McGuinness 2001) to construct the food ontology,

which contains nutritional facts for each food such as

amounts of nutrients, calories, and portion size. (4)

Brief personal profiles of users, containing such data as

sex, age, height, and weight, are also built into the

personal profile ontology by domain experts. (5)

Domain experts use the FML editor to build the OMAS

knowledge base and rule base.

• Utilization Phase: The OMAS performs the following

actions. (1) Find a user’s personal profile. (2) Acquire

necessary dietary knowledge by analyzing nutritional

facts for foods eaten, calculate the percentage of

calories from nutrients, and identify the planned dietary

goal. (3) Infer healthy diet status. (4) Generate semantic

sentences and output them to the user. (5) The

generated semantic descriptions for foods eaten are

stored in the healthy diet status repository. (6) The

healthcare domain experts evaluate the performance of

the proposed approach. (7) Last, providing the output as

a reference for users to establish a healthy lifestyle.

4.2 Personal knowledge agent

The personal knowledge agent plays a role in the retrieval of

personal profile data, such as age, sex, height, and weight,

Table 3 Harris–Benedict Equation

Sex Basal metabolic rate (BMR)

Male 66 ? (13.7 9 Weight) ? (5 9 Height) - (6.8 9 Age)

Female 655 ? (9.6 9 Weight) ? (1.8 9 Height) - (4.7 9 Age)


123

Table 4 Algorithm of the

personal knowledge agentPersonal knowledge agent algorithm Input: 1. Input the ontology repository, including food ontology, personal profile ontology, and FML-based knowledge

base and rule base. 2. Input the foods eaten set M

{[ N1 MainCourseMainCourse ,..., ],

[ Q1 SideDishSideDish ,..., ],

[ R1 DessertDessert ,..., ],

[ L1 BeverageBeverage ,..., ]}

/*where (1) N, Q, R, and L denote the number of food eaten grouped into main course, side dish, dessert, and beverage, respectively. (2) Foods eaten means that the eaten food that the user inputs via the OMAS platform.*/

3. Input the portions of each food eaten set P

{[ N1 MainCoursePortionMainCoursePortion _,...,_ ],

[ Q1 SideDishPortionSideDishPortion _,...,_ ],

[ R1 DessertPortionDessertPortion _,...,_ ],

[ L1 BeveragePortionBeveragePortion _,...,_ ]}

/*where the portions of each food eaten means the portions of the food that the user eats. */ Output:

AgeSet , BMISet , PCCSet , PCPSet , PCFSet , and CDSet

Method: Step1: If user’s identification is passed Then

Step1.1: Retrieve the user’s Age ( Value Age), Height(Value Height), and Weight(Value Weight) from the ontology

repository to calculate the user’s BMI by Value BMI2

HeightWeight ValueValue

Step1.2: Add Value Age to AgeSet and add Value BMI to BMISet

Step2: Retrieve nutrition facts (calories per portion, the number of carbohydrate, protein, and fat grams contained in one portion) of eaten main courses, side dishes, desserts, and beverages from the ontology repository.

C{[ N1 MainCourseCMainCourseC _,...,_ ],

[ Q1 SideDishCSideDishC _,...,_ ],

[ R1 DessertCDessertC _,...,_ ],

[ L1 BeverageCBeverageC _,...,_ ]}

teCarbohydraG

{[ NteCarbohydra1teCarbohydra MainCourseGMainCourseG _,...,_ ],

[ QteCarbohydra1teCarbohydra SideDishGSideDishG _,...,_ ],

[ RteCarbohydra1teCarbohydra DessertGDessertG _,...,_ ],

[ LteCarbohydra1teCarbohydra BeverageGBeverageG _,...,_ ]}

ProteinG

{[ NProtein1Protein MainCourseGMainCourseG _,...,_ ],

[ QProtein1Protein SideDishGSideDishG _,...,_ ],

[ RProtein1Protein DessertGDessertG _,...,_ ],

[ LProtein1Protein BeverageGBeverageG _,...,_ ]}

FatG

{[ NFat1Fat MainCourseGMainCourseG _,...,_ ],

[ QFat1Fat SideDishGSideDishG _,...,_ ],

[ RFat1Fat DessertGDessertG _,...,_ ],

[ LFat1Fat BeverageGBeverageG _,...,_ ]}

Step3: Calculate the total calories intake, the total calories from carbohydrate, protein, and fat, respectively, by

Step3.1: ∑ × PCC Actual

Step3.2: 4)( ××∑ PGC teCarbohydrateCarbohydra

Step3.3: 4)( ××∑ PGC ProteinProtein

Step3.4: 9)( ××∑ PGC FatFat

Step4: Calculate the values of PCC, PCP, and PCF by

Step4.1: ValuePCC ( teCarbohydraC / ActualC ) × 100% and add Value PCC to PCCSet

Step4.2: Value PCP ( ProteinC / ActualC ) × 100% and add Value PCP to PCPSet

Step4.3: ValuePCF ( FatC / ActualC ) × 100% and add Value PCF to PCFSet

Step5: Retrieve the planned calories needs ( PlannedC ) from the ontology repository

Setp6: Calculate the value of CD by

Value CD | PlannedC - ActualC | and add Value CD to CDSet

Setp7: End


123

from the personal profile ontology. With this information,

the personal knowledge agent can determine user’s BMI and

the planned daily caloric intake using the Harris–Benedict

equation (Frankenfield et al. 1998) (Table 3).

Additionally, the personal knowledge agent further

examines the amount of carbohydrate, protein, and fat

grams for one portion in the collected meal records using

the food ontology. The number of calories contained in one

Age

Age

1

20 30 40

Young Middle Old

BMI (kg/m^2)

BMI

1

Under-Weight

PCC (%)

PCC

0

1

PCP (%)0

PCP

1

PCF (%)0

PCF

1

CD (kcal)0

CD

1

Normal OverWeight

Low Balanced High Low Balanced High

Low Balanced High Acceptable More-or-LessUnAcceptable UnAcceptable

HDS0

HDS

1

Very-UnHealthy

UnHealthy Medium-Healthy

Healthy VeryHealthy

50 60 15 20 25 30 35 40

10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100

10 20 30 40 50 60 70 80 90 100 25 50 75 100 125 150 175 200 225 250

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

(g)

(e)

(c)

(a) (b)

(d)

(f)

Fig. 8 Trapezoidal membership functions for fuzzy variables a Age, b BMI, c PCC, d PCP, e PCF, f CD, and g HDS


123

portion is also acquired. With these nutritional facts for

foods consumed, the personal knowledge agent transforms

these facts into actual calories, percentage of calories from

carbohydrate, percentage of calories from protein, and

percentage of calories from fat. Finally, the personal

knowledge agent retrieves planned daily caloric intake to

calculate the caloric difference between actual calories and

planned caloric intake. Table 4 lists the algorithm of the

personal knowledge agent.

4.3 Fuzzy inference agent

The main task of the fuzzy inference agent is to infer the

level of a user achieving a healthy dietary state according

to the constructed knowledge base and rule base of the

OMAS. First, five input fuzzy variables, Age, BMI, PCC,

PCP, PCF, and CD, as well as one output fuzzy variable,

HDS, are considered. In this paper, a trapezoidal mem-

bership function for a fuzzy set (FS), denoted by Eq. 1, is

Table 5 Algorithm of the fuzzyinference agent

Fuzzy Inference Agent Algorithm Input: 1. Input the ontology repository, including food ontology, personal profile ontology, and FML-based knowledge base and rule base. 2. AgeSet , BMISet , PCCSet , PCPSet , PCFSet , and CDSet

Output: Healthy diet status set HDSSet

Method: Step1: Retrieve the knowledge base of the OMAS FML from the ontology repository and parse it to get the input fuzzy set ( inA ) and output fuzzy set ( outA )

inA ={ AgeYoung, AgeMiddle, AgeOld, BMILow, BMIMedium, BMIHigh, PCCLow, PCCBalanced, PCCHigh, PCPLow,

PCPBalanced, PCPHigh, PCFLow, PCFBalanced, PCFHigh, CDAcceptable, CDMore-or-LessUnAcceptable, CDUnAcceptable}outA ={HDSVeryUnHealthy, HDSUnHealthy, HDSMediumHealthy, HDSHealthy, HDSVeryHealthy}

Step2: Retrieve the rule base of the OMAS FML from the ontology repository and parse it to get the fuzzy inference rules set FIR = {Rule1, Rule2, …, Rule K}Step3: Implement the fuzzy inference

Let )],,,,,[( CDPCFPCPPCCBMIAge ValueValueValueValueValueValueX

Step3.1: For k 1 to K /*K denotes the number of fuzzy rules*/ Step3.1.1: Calculate the matching degree of the Rule k by

))(( XMINinAk μμ

Step3.1.2: Calculate the center of area of the Rule k by

)(_ _ kout COAkAy μ

Step3.2: For t 1 to T /*T means the number of linguistic terms of the output fuzzy variable, HDS, that is, T = 5 in this paper.*/Step3.2.1: Calculate the membership values of X to the fuzzy classes tF by )( kAt out

yMAXy

/*where (1) Fuzzy classes mean the linguistic terms of the output fuzzy variable (HDS), that is, VeryUnHealthy , UnHealthy, MediumHealthy, Healthy, and VeryHealthy in this paper, and (2) tF

means the fuzzy class for all of fuzzy rules and each fuzzy class is an aggregation of the fired rules that have the same consequences.*/

Step3.2.2: Defuzzify into a crisp value by

∑∑==

T

tt

T

tttHDS wywValue

11

/*where tw means the weight for ty */

Step3.2.3: Add HDSValue to HDSSet

Setp4: End

Table 6 Sentence patterns Semantic Analysis Sentence:The eaten items at meal by this user exhibit that the person is at [FNAge: Young , Middle, Old] age and the

body mass index is [FNBMI: UnderWeight, Normal, OverWeight ], meanwhile percentage of calories from carbohydrate is [FNPCC: Low, Balanced, High], percentage of calories from protein is [FNPCP: Low, Balanced, High], percentage of calories from fat is [FNPCF: Low, Balanced, High], and caloric difference is [FNCD: Acceptable, More-or-LessUnAcceptable, UnAcceptable].Semantic Decision Sentence:

The OMAS justifies that the level of healthy diet status for meal is [FNHDS: VeryUnHealthy, UnHealthy, MediumHealthy, Healthy, VeryHealthy ]. (Level: [0, 1])


123

used by the domain experts according to the past heuristic

experience for simplifying computational complexity.

The FS is specified by four parameters FS(x: param1,

param2, param3, param4) and can be expressed as

[param1, param2, param3, param4] (Lee and Wang 2008).

Another reason for adopting the trapezoidal membership

function in this paper is that when param2 equals param3,

then the trapezoidal membership function will be reduced

to the triangular membership function. Fuzzy variable Age

has three linguistic terms, AgeYoung, AgeMiddle, and AgeOld,

that express user’s age, whose trapezoidal membership

functions are [20, 20, 25, 30], [25, 33, 38, 50], and [45, 50,

60, 60], respectively. According to the BMI definitions (1)

Underweight, BMI \ 18.5; (2) Normal, 18.5 B BMI \ 24;

(3) Overweight, 24 B BMI \ 30; (4) Obese, BMI C 30,

the BMI fuzzy variable defines three linguistic terms,

namely, BMIUnderWeight, BMINormal, and BMIOverWeight,

whose trapezoidal membership functions are [15, 15, 18.5,

20], [18.5, 20, 22, 24], and [22, 24, 40, 40], respectively.

The suggested percentages of calories from carbohy-

drate, protein, and fat are 55–65, 10–20, and 25–35%,

respectively. Hence, the linguistic terms of fuzzy variable

Table 7 Algorithm of the

semantic generation agentSemantic Generation Agent Algorithm Input: 1. Input the ontology repository2. Healthy diet status set HDSSet

3. Input fuzzy set inA and output fuzz set outA

inA ={ AgeYoung, AgeMiddle, AgeOld, BMILow, BMIMedium, BMIHigh, PCCLow, PCCBalanced, PCCHigh, PCPLow,

PCPBalanced, PCPHigh, PCFLow, PCFBalanced, PCFHigh, CDAcceptable, CDMore-or-LessUnAcceptable, CDUnAcceptable}outA ={HDSVeryUnHealthy, HDSUnHealthy, HDSMediumHealthy, HDSHealthy, HDSVeryHealthy}

Output: Semantic analysis for the eaten meal Method: Step1: Let )],,,,,[( CDPCFPCPPCCBMIAge ValueValueValueValueValueValueX

Step1.1: ))(),(),(( AgeAgeAgeAgeAgeAgeAge ValueValueValueMAXOldMiddleYoung

μμμμ and save the

linguistic term to AgeLT

Step1.2: ))(),(),(( BMIBMIBMIBMIBMIBMIBMI ValueValueValueMAXOverweightNormaltUnderweigh

μμμμ and save

the linguistic term to BMILT

Step1.3: ))(),(),(( PCCPCCPCCPCCPCCPCCPCC ValueValueValueMAXHighBalancedLow


linguistic term to PCCLT

Step1.4: ))(),(),(( PCPPCPPCPPCPPCPPCPPCP ValueValueValueMAXHighBalancedLow


linguistic term to PCPLT

Step1.5: ))(),(),(( PCFPCFPCFPCFPCFPCFPCF ValueValueValueMAXHighBalancedLow


linguistic term to PCFLT

Step1.6: ))(),(),(( CDCDCDCDCDCDCD ValueValueValueMAXleUnAcceptabptableLessUnAcceorMoreAcceptable

μμμμ−−

and

save the linguistic term to CDLT

Step1.7:))(),(

),(),(),((

HDSHDSHDSHDS

HDSHDSHDSHDSHDSHDSHDS

ValueValue

ValueValueValueMAX

yVeryHealthHealthy

thyMediumHealUnHealthythyVeryUnHeal

μμ

μμμμ

and save the linguistic term to HDSLT

Step2: Output meal’s healthy state using semantic sentences The eaten items at meal by this user exhibit that the person is at AgeLT age and the body mass index

is BMILT , meanwhile percentage of calories from carbohydrate is PCCLT , percentage of calories from

protein is PCPLT , percentage of calories from fat is PCFLT , and caloric difference is CDLT

The OMAS justifies that the level of healthy diet state for meal is HDSLT (Level: HDSμ )

Step3: End

FS x : param1; param2; param3; param4ð Þ ¼

0 x\param1

x� param1ð Þ= param2� param1ð Þ param1� x\param2

1 param2� x� param3

param4� xð Þ= param4� param3ð Þ param3\x� param4

0 x [ param4

8>>>><

>>>>:

ð1Þ


123

PCC are PCCLow, PCCBalanced, and PCCHigh, whose trap-

ezoidal membership functions are [0, 0, 50, 55], [50, 55,

65, 70], and [65, 70, 100, 100], respectively. The linguistic

terms of fuzzy variable PCP are PCPLow, PCPBalanced, and

PCPHigh, whose trapezoidal membership functions are [0,

0, 5, 10], [5, 10, 20, 25], and [20, 25, 100, 100], respec-

tively. The linguistic terms of fuzzy variable PCF are

PCFLow, PCFBalanced, and PCFHigh, whose trapezoidal

membership functions are [0, 0, 20, 25], [20, 25, 35, 40],

and [35, 40, 100, 100], respectively. The membership

functions of fuzzy variable CD are CDAcceptable, CDMore-or-

LessUnacceptable, and CDUnacceptable, whose trapezoidal

membership functions are [0, 0, 50, 100], [70, 100, 150,

200], and [150, 200, 5000, 5000], respectively.

Second, the fuzzy inference mechanism performs

membership functions to compute the membership degrees

for each meal recorded in the ontology repository. The

MIN operator is then used to combine the degree of match

between each fuzzy rule’s conditions. Third, the area center

of each rule is calculated. The HDS fuzzy variable is uti-

lized with five linguistic terms to represent the level of

healthy dietary state for the meal eaten. The linguistic

terms are HDSVeryUnHealthy, HDSUnHealthy, HDSMediumHealthy,

HDSHealthy, and HDSVeryHealthy, whose trapezoidal mem-

bership functions are [0, 0, 0.1, 0.25], [0.1, 0.25, 0.25, 0.5],

[0.25, 0.5, 0.5, 0.75], [0.5, 0.75, 0.75, 0.9], and [0.75, 0.9,

1.0, 1.0], respectively. Fourth, the fuzzy inference mecha-

nism performs the MAX operation to integrate triggered

rules with the same consequences and outputs the maxi-

mum area center. The parameters of all membership

functions are determined by domain experts. Figures 8a–g

show the trapezoidal membership functions for fuzzy

variables Age, BMI, PCC, PCP, PCF, CD, and HDS,

respectively. Finally, the level of a healthy dietary status of

a meal eaten is inferred. Table 5 shows the fuzzy inference

agent algorithm.

4.4 Semantic generation agent

According to sentence patterns (Table 6), the semantic

generation agent retrieves the inferred results obtained by

the fuzzy inference agent and then transforms these inferred

results into knowledge to present healthy dietary status

using semantic descriptions.

Table 7 shows the semantic generation agent algorithm.

For example, if a 23-year-old woman eats two portions of

strawberry chocolate and drinks one portion of Coca-Cola,

the OMAS indicates that the level of a healthy dietary state

for such a meal is very unhealthy when the person is young

and their body mass is underweight, the percentage of

calories from carbohydrates is high, the percentage of

calories from protein is low, the percentage of calories

from fat is low, and the caloric difference is unacceptable.

5 Experimental results

The OMAS platform was implemented by ASP.Net (Web

Inference), Microsoft C# programming language (OMAS),

and Java (FML and Visual Editor). This paper focuses on

people aged 20–60 years old and with average levels of

Table 8 Personal information of the involved people

No. Sex Age Height (cm) Weight (kg)

1 Male 30 170 60

2 Male 24 169 71

3 Male 24 166 73

4 Male 25 172 85

5 Male 23 171 68

6 Male 26 178 81

7 Male 24 165 65

8 Male 27 173 75

9 Male 31 167 70

10 Male 28 181 88

11 Female 23 156 50

12 Female 23 161 51

13 Female 24 170 44

14 Female 24 168 48

15 Female 25 170 52

16 Female 25 156 46

17 Female 24 163 52

18 Female 26 155 54

19 Female 27 158 58

20 Female 28 165 51

Table 9 Part of the meal records of the subject 1

Meal no. Main courses (portion) Side dishes (portion) Desserts (portion) Beverages (portion)

1 Hawaii pizza (1) Composited fruits (1) Egg pudding (1) Green tea (1)

9 Royal lunch box vinegar chicken (1) Grapes toast (1) Milk Chocos (1) Seattle black coffee (1)

10 Glutinous oil rice (1) Composited fruits (1) Vanilla chocolate Bar (1) Mineral water (1)

20 American smoked sausage burger (1) Gratin cheese potato (1) Strawberry chocolate (1) Oat milk (1)


123

physical activity. Twenty students attending National

University of Tainan (NUTN), Taiwan, were enrolled in

this experiment. These subjects recorded their dinner meals

on Monday–Friday for about 1 month. Therefore, 20 die-

tary records were collected for each subject for a total of

400 dietary records. Table 8 lists the personal information

of these subjects. There are two kinds of domain experts,

including ontology domain experts and healthcare domain

experts, involved in those following experiments shown in

this section. The healthcare domain experts did those

experimental evaluations and the ontology domain experts

built the fuzzy rules together.

5.1 Evaluation results of OMAS

Table 9 lists meals 1, 9, 10, and 20 eaten by subject 1. In

Fig. 9, the OMAS indicates that the level of a healthy

dietary state for the meal 10 eaten by subject 1 is

VeryHealthy.

Table 10a lists the OMAS outcomes for meals 1, 9, 10,

and 20 of subject 1; these outcomes were VeryUnHealthy,

VeryUnHealthy, VeryHealthy, and Healthy, respectively.

This OMAS result for meal 10 is exactly the same as that

generated by the domain expert, meaning that both the

domain expert and the OMAS categorized this meal as

VeryHealthy. Therefore, eating meal 10 is a healthy choice

for subject 1. However, subject 1 should not eat unhealthy

meal 9. In other words, the proposed OMAS can inform

subject 1 about which foods he should avoid and which

foods he should eat. Additionally, the experimental results

of the subject 12 are also described. Table 10b lists the

outcomes of meals 3, 5, 8 and 16 eaten by subject 12.

Table 11 lists some of collected meal records for subject

12.

5.2 Analysis of OMAS

The second experiment is still to consider subject 1 and

subject 12 as examples to assess the variance in the level of

healthy dietary status evaluated by the domain expert and

the OMAS.

Figures 10a, b show the charts of the level of a healthy

dietary status evaluated by the domain expert and the

OMAS for each meal record for subject 1 and subject 12,

respectively. Figure 10a reveals that the difference in

generated level of a healthy dietary status by the domain

expert and the OMAS for meals 1–20 is \0.25, excluding

meal 1. Figure 10b indicates that the difference between

the level of healthy dietary status by the domain expert and

the OMAS was \0.2 for all meal records.

5.3 Accuracy, precisions, and recall for OMAS

The third experiment evaluates the performance of the

accuracy, precision, and recall for the OMAS. Table 12

shows four possible outcomes of a single prediction (Lee

and Wang 2007). The accuracy, precision, and recall are

calculated by Eqs. 2, 3 and 4, respectively. In this

Fig. 9 Screenshot of meal 10 of the subject 1


123

Table 10 Outcomes of part of meals of the subject 1 and subject 12

Age BMI PCC PCP PCF CD

Subject 1

Meal 1 30 20.76 56.71 7.56 35.73 156.22

The eaten items at meal by this user exhibit that the person is at Middle age and the body mass index is Normal, meanwhile percentage

of calories from carbohydrate is Balanced, percentage of calories from protein is Low, percentage of calories from fat is Balanced,

and caloric difference is More-or-LessUnAcceptable

The OMAS justifies that the level of healthy diet state for meal is VeryUnHealthy. (Level: 0.37)

The domain expert justifies that the meal is Healthy

Meal 9 30 20.76 58.85 11.07 30.08 1150.82


of calories from carbohydrate is Balanced, percentage of calories from protein is Balanced, percentage of calories from fat is

Balanced, and caloric difference is UnAcceptable


The domain expert justifies that the meal is VeryUnHealthy

Meal 10 30 20.76 64.5 10.75 24.75 48.22


of calories from carbohydrate is Balanced, percentage of calories from protein is Balanced, percentage of calories from fat is Low,

and caloric difference is Acceptable

The OMAS justifies that the level of healthy diet state for meal is VeryHealthy. (Level: 0.9)

The domain expert justifies that the meal is VeryHealthy

Meal 20 30 20.76 55 15 30 135.62


of calories from carbohydrate is Balanced, percentage of calories from protein is Balanced, percentage of calories from fat is

Balanced, and caloric difference is More-or-LessUnAcceptable

The OMAS justifies that the level of healthy diet state for meal is Healthy. (Level: 0.71)


Subject 12

Meal 3 23 19.68 58.82 7.97 33.22 1443.84

The eaten items at meal by this user exhibit that the person is at Young age and the body mass index is UnderWeight, meanwhile

percentage of calories from carbohydrate is Balanced, percentage of calories from protein is Low, percentage of calories from fat is




Meal 5 23 19.68 57.66 7.21 35.14 69.84



Balanced, and caloric difference is Acceptable

The OMAS justifies that the level of healthy diet state for meal is UnHealthy. (Level: 0.59)


Meal 8 23 19.68 76.03 1.65 22.31 220.84


percentage of calories from carbohydrate is High, percentage of calories from protein is Low, percentage of calories from fat is Low,and caloric difference is UnAcceptable



Meal 16 23 19.68 56.48 9.97 33.55 820.44







123

experiment, the healthcare domain experts first classify

the food eaten into two categories: ‘‘healthy’’ and

‘‘unhealthy.’’ Next, based on various thresholds in the

interval of [0.05, 0.90], the OMAS also classifies the

healthy diet state of the food eaten. Finally, the curves

for accuracy, precision, and recall are acquired according

to Table 12. Figures 11a, b show the curves of the

accuracy, precision, and recall for subject 1 and subject

12, respectively. Herein, a threshold denotes the mem-

bership degree threshold for HDS. Figure 11a reveals

that accuracy is 90% when the threshold is 0.55–0.70

and Fig. 11b shows that accuracy was [80% for most

thresholds. Additionally, Fig. 11 also indicates that recall

and precision are inversely related; that is, as recall

increases, precision decreases.

Table 11 Part of the meal records of the subject 12

Meal no. Main courses (portion) Side dishes (portion) Desserts (portion) Beverages (portion)

3 Gratin curry chicken rice (1) Milk bread with pudding (2) Vanilla chocolate bar (1) Yakult fermented milk (2)

5 – Japanese style corn soup (1) Strawberry wafer roll (1) –

8 – – Strawberry chocolate (2) Coca Cola (1)

16 Rice roll with tempura shrimp (2) Composited fruits (1) Milk Chocos (1) –

Fig. 10 Charts of the level of

the healthy diet status for asubject 1 and b subject 12

Table 12 Classification of results

Actual class Predicted class

Yes No

Yes True positive (TP) False negative (FN)

No False positive (FP) True negative (TN)


123

Accuracy ¼ TN þ TPð Þ= TN þ FN þ FPþ TPð Þ � 100%

ð2ÞPrecision ¼ TP= TPþ FPð Þ � 100% ð3ÞRecall ¼ TP= TPþ FNð Þ � 100% ð4Þ

5.4 Performance of OMAS

The fourth experiment evaluates the mean square error

(MSE), calculated by Eq. 5, between outcomes of the

domain expert and the OMAS. Figure 12 lists the MSE of

each subject for all meal records; all MSEs are \0.5.

MSE ¼ 1

n

Xn

i¼1

e2i ð5Þ

where n is the total number of subjects, and ei is the out-

come difference between the OMAS and domain expert.

6 Conclusions

While living a healthy life is important, many individuals

have difficulty in determining whether what they eat is

healthy. Therefore, based on the physical characteristics of

an individual and opinions about diet, the OMAS is pro-

posed to represent semantic descriptions such that indi-

viduals can understand whether the food they eat is

healthy. The proposed OMAS contains three sub-agents, a

personal knowledge agent, a fuzzy inference agent, and a

semantic generation agent. Experimental results indicate

that the proposed system enables the intelligent behavior

for a suitable healthy diet for users based on the con-

structed ontology. The proposed agent helps people eat

healthily and keep their bodies in shape, and reduces the

workload of medical experts. Additionally, according to

the viewpoint of the nutritionists, the food group balance is

Fig. 11 Curves of the

accuracy, precision, and recall

for a subject 1 and b subject 12

Fig. 12 MSE of the each

subject


123

also a key factor to evaluate whether the food eaten is

healthy or not. Therefore, adding the food group balance to

the proposed OMSA will be one of the future research

works. Additionally, the concept of a type-2 fuzzy set and a

learning mechanism for follow-up research will be also

considered in the future.

Acknowledgments This work is supported by the National Science

Council (NSC) of Taiwan under the grant NSC97-2221-E-024-011-

MY2 and NSC98-2221-E-024-009-MY3. The authors would like to

thank all subjects for their involving this research project. Ted Knoy

is appreciated for his editorial assistance.

References

Acampora G, Loia V (2005a) Using FML and fuzzy technology in

adaptive ambient intelligence environments. Int J Comput Intell

Res 1(2):171–182

Acampora G, Loia V (2005b) Fuzzy control interoperability and

scalability for adaptive domotic framework. IEEE Trans Ind

Inform 1(2):97–111

Caceres C, Fernandez A, Ossowski S, Vasirani M (2006) Agent-based

semantic service discovery for healthcare: an organizational

approach. IEEE Intell Syst 21(6):11–20

Chan V, Ray P, Parameswaran N (2008) Mobile e-Health monitoring:

an agent-based approach. IET Commun 2(2):223–230

Chen RS, Chen DK (2008) Apply ontology and agent technology to

construct virtual observatory. Exp Syst Appl 34(3):2019–2028

Debenham J, Sierra C (2008) Merging intelligent agency and the

Semantic Web. Knowledge Based Syst Arch 21(3):184–191

Ferber J (1999) Multi-Agent Systems: an introduction to distributed

artificial intelligence. Addison-Wesley, New York

Frankenfield DC, Muth ER, Rowe WA (1998) The Harris-Benedict

studies of human basal metabolism: history and limitations.

J Am Diet Assoc 98(4):439–445

Lee CS, Wang MH (2007) Ontology-based intelligent healthcare

agent and its application to respiratory waveform recognition.

Exp Syst Appl 33(3):606–619

Lee CS, Wang MH (2008) Ontological fuzzy agent for electrocar-

diogram application. Exp Syst Appl 35(3):1223–1236

Lee CS, Jian ZW, Huang LK (2005) A fuzzy ontology and its

application to news summarization. IEEE Trans Syst Man

Cybern B 35(5):859–880

Lee CS, Jiang CC, Hsieh TC (2006) A genetic agent using ontology

model for meeting scheduling system. Inf Sci 176(9):131–1155

Lee CS, Wang MH, Acampora G, Loia V, Hsu CY (2009) Ontology-

based intelligent fuzzy agent for diabetes application. In:

Proceeding of the 2009 IEEE Symposium on Computational

Intelligence for Intelligent Agents (IA 2009), Nashville, Ten-

nessee, USA, March 30–April 2, pp 16–22

Moreno A (2006) On the evolution of applying agent technology to

healthcare. IEEE Intell Syst 21(6):8–10

Moreno A, Valls A, Isern D, Sanchez D (2006) Applying agent

technology to healthcare: the CruSMA experience. IEEE Intell

Syst 21(6):63–67

Noy NF, McGuinness DL (2001) Ontology development 101: a guide

to creating your first ontology. Stanford knowledge systems

laboratory technical report KSL-01-05 and Stanford medical

informatics technical report SMI-2001-0880

Orgun B, Vu J (2006) HL7 ontology and mobile agents for

interoperability in heterogeneous medical information systems.

Comput Biol Med 36(7–8):817–836

Reformat M, Ly C (2009) Ontological approach to development of

computing with words based systems. Int J Approx Reason

50(1):72–91

Sanchez FG, Garcı́a RV, Bejar RM, F. Breis JT (2009) An ontology,

intelligent agent-based framework for the provision of semantic

web services. Exp Syst Appl 36(2):3167–3187

US Food and Drug Administration (2009) How to understand and use

the nutrition facts label. http://www.cfsan.fda.gov/*dms/

foodlab.html. Accessed 2 July 2009

Wang MH, Lee CS, Hsieh KL, Hsu CY, and Chang CC (2009)

Intelligent ontological multi-agent for healthy diet planning. In:

Proceeding of the 2009 IEEE International Conference on Fuzzy

System (FUZZ-IEEE 2009), Jeju Island, Korea, Aug. 19–24, pp

751–756

Zunino A, Campo M (2009) Chronos: a multi-agent system for

distributed automatic meeting scheduling. Exp Syst Appl

36(3):7011–7018


123

http://www.cfsan.fda.gov/~dms/foodlab.html

http://www.cfsan.fda.gov/~dms/foodlab.html

Ontology-based multi-agents for intelligent healthcare applications

Documents

Transcript of Ontology-based multi-agents for intelligent healthcare applications