An Embedded Mobile ECG Reasoning System for Elderly Patients

An Embedded Mobile ECG Reasoning System for

Elderly Patients

Dong-Her Shih, Hsiu-Sen Chiang, Binshan Lin, Member, IEEE, and Shih-Bin Lin

Abstract—With the increase in the number senior citizens and chronic diseases, the number of elderly patients who need con-stant assistance has increased. One key point of all critical care for elderly patient is the continuous monitoring of their vital signs. Among these, the ECG signal is used for noninvasive diagnosis of cardiovascular diseases. Also, there is a pressing need to have a proper system in place for patient identification. Errors in patient identification, and hence improper administration of medication can lead to disastrous results. This paper proposes a novel em-bedded mobile ECG reasoning system that integrates ECG signal reasoning and RF identification together to monitor an elderly patient. As a result, our proposed method has a good accuracy in heart beat recognition, and enables continuous monitoring and identification of the elderly patient when alone. Moreover, in order to examine and validate our proposed system, we propose a man-agerial research model to test whether it can be implemented in a medical organization. The results prove that the mobility, usabil-ity, and performance of our proposed system have impacts on the user’s attitude, and there is a significant positive relation between the user’s attitude and the intent to use our proposed system.

Index Terms—ECG, embedded system, fuzzy Petri net (FPN), RF identification (RFID), telemedicine.

I. INTRODUCTION

T

HE DEMOGRAPHIC shift of an aging population is a global phenomenon. Population ageing slows work forces and economic growth, and as health expenditure is the highest for the elderly, it puts increased demands on health and aged care systems [1]. In general, approximately 75% of the elderly have one or more chronic diseases, among which the cardiovascular disease is one of the most common diseases. Cardiovascular diseases are the greatest and serious health threat in an aging society [2]. The patients need more care from their family or friends. Unfortunately, more than 27% of the country’s elderly population currently live alone [3]. Therefore, elderly health management is an important issue, especially with regards to cardiovascular disease.

Manuscript received December 12, 2006; revised March 31, 2008 and June 6, 2008. First published April 28, 2009; current version published June 3, 2010. This work was supported in part by the National Science Council of Taiwan under Grant NSC95-3113-P-224-004, Grant NSC97-2218-E-224-003, Grant NSC95-2221-E-224-049-MY2, Grant NSC 95-3113-P-224-004, and Grant NSC97-2218-E-224-003.

D.-H. Shih is with the Department of Information Management, National Yunlin University of Science and Technology, Douliu 64002, Taiwan (e-mail: shihdh@yuntech.edu.tw).

H.-S. Chiang is with the Department of Information Management, Da-Yeh University, Changhua 51591, Taiwan (e-mail: chianghs@mail.dyu.edu.tw).

B. Lin is with the College of Business Administration, Louisiana State Uni-versity, Shreveport, LA 71115 USA (e-mail: binshan.lin@lsus.edu).

S.-B. Lin is with the Department of Information Management, National Chiao Tung University, Hsinchu 300, Taiwan (e-mail: g9423701@yuntech.edu.tw).

Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TITB.2009.2021065

In the past, most of the patient monitoring devices were de-signed for hospital. Actually, a person may suffer from car-diovascular disease at any time and any place. Advanced de-velopments in existing information technologies are making it possible to extend the range of treatment and health services available at the patient’s home [4]. This kind of health care sys-tem can carefully monitor the health condition of the patient for a long time and within their familiar environment.

In most healthcare systems, the vital signals are widely used in the diagnosis and therapy of many diseases. Among these, the ECG signal is used for noninvasive diagnosis of cardiovascular diseases [5]. In the past decades, many successful healthcare applications of ECG have been presented. However, these cur-rently available systems have some drawbacks [6], [7].

First, some developments of the healthcare system still focus on monitoring aspect and lack the focus on warning aspect. In case of an emergency, there are no predefined procedures that will help the patient quickly. Thus, it is extremely important to study new healthcare systems to achieve a true meaningful mon-itoring and alert system [8]. Second, sometimes the art devices for healthcare system are quite clumsy and uncomfortable for long-term use. Besides, most of them are usually designed for clinical professionals, losing the end-user viewpoint [9]. Third, there is a lack of real-time patient identification device. Accord-ing to a report [10], more than 770 000 patients die or suffer injuries each year while in hospital due to medical errors such as the incorrect medication dosage. Aggregated across the U.S., the total cost of these mishaps is between $1.5 and $5 billion annually. This is mainly due to the lack of an efficient way to identify the patient being treated [11]. Therefore, a real-time patient identification technology plays an important role in the healthcare system.

In order to overcome all these drawbacks, we proposed a new mobile ECG monitoring and alert system architecture that integrates current network technology such as general packet radio service (GPRS), embedded system, and RF identifica-tion (RFID) technologies. When our proposed system detects an abnormal ECG activity, related hospitals will receive the alarm message. The prototype of our proposed mobile ECG monitoring and alert system enables continuous monitoring and identification of the elderly patient who lives alone.

II. BACKGROUND ANDRELATEDWORKS

An ECG is a graphic produced by an electrocardiograph, which records the electrical voltage in the heart in the form of a continuous strip graph. It is the prime tool in cardiac elec-trophysiology, and plays a vital role in screening and diagnosis of cardiovascular diseases. A single normal cycle of the ECG 1089-7771/$26.00 © 2009 IEEE

represents the successive arterial depolarization and ventricu-lar repoventricu-larization, which occurs with every heartbeat. These can be approximately associated with the peaks and other ECG waveforms, which are labeled as P, Q, R, S, and T [12]. The P wave represents the both atria activations (depolarization). The Q wave represents early depolarization that causes a small downward deflection after the P wave. The R wave is a large deflection of the ECG when the ventricles are depolarized. The S wave is the first wave after the R wave that dips below the baseline. The T wave represents the ventricular repolarization. Recently, different characteristics, such as PR, QRS, and ST intervals, have also been used in diagnosis [13]. The QRS com-plex represents the ventricular activation. The PR interval is between the beginning of the P wave and the beginning of the QRS complex. The section from the end of the QRS complex till the end of a T wave is known as the ST interval.

Automated classification of heartbeats has been previously reported by other investigators by using a variety of features to represent the ECG and a number of classification methods. These features include RR interval, heartbeat interval, ECG morphology [14], ST segment deviation, ST segment slope, ST segment area (STA), T-normal amplitude [15], STA, R-S in-terval (RSI), ST slope (STS), R-T inin-terval (RTI), QRS area (QRSA), Q-T interval (QTI), R-wave amplitude (RWA), heart beat rate (HBR), statistical features QRS energy (QRSE), mean of the power spectral density (MPSD), autocorrelation coeffi-cient (ACC), signal histogram (SH) [16], discrete Fourier trans-form (DFT) coefficients [17], statistical features of the QRS complexes [18], Hermite coefficients [19], shift invariant [20], continuous wavelet transform coefficients [21], QRS complex wave width [22], amplitude value, discrete cosine transform (DCT) coefficients, and discrete wavelet transform (DWT) co-efficients [23]–[25].

In addition, the classification of heartbeats is an important step toward identifying an arrhythmia. Classification methods em-ployed include linear discriminants [18], association rules [15], neural network [16], [17], [25], fuzzy neural network [18], [19], [20], [24], hidden Markov models [21], mirrored Gauss model [22], artificial immune recognition system [26], and sup-port vector machine [23]. Some widely used methods and fea-tures for classification are listed in Table I.

In traditional rule-based reasoning methods [27], [28], the causality relationship is described as a number of rules. How-ever, since many rules are involved, a powerful inference engine is usually required to avoid cyclical causality to discover the con-nection between rules and derive the final decision. Obviously, such a rule-based method is time- and memory-consuming, and lacks self-learning ability [29]. Recently, some researchers adopt Petri nets (PNs) for inference and reasoning. PNs have the ability to represent and analyze a rule-based system in con-current and synchronous phenomena. PNs also have an inherent quality in representing logic in an intuitive and visual way, and fuzzy PNs (FPNs) have all the advantages of PNs. Many results prove that an FPN is suitable for representing and reasoning misty logic implication relations [30]. Therefore, we use FPN as our reasoning process in our proposed embedded mobile ECG reasoning system.

TABLE I

SOMEUSEDMETHODS ANDFEATURES FORCLASSIFICATION

Many ECG-monitoring systems using mobile technologies for patients have been proposed. Hung and Zhang [31] present a wireless application protocol (WAP) based telemedicine system to monitor a patient. It utilizes WAP devices as mobile access terminals for general inquiry and patient-monitoring services. A wireless ECG subsystem was built for recording ambulatory ECG in an indoor environment and for storing ECG data into the database. The system shows how feasible WAP can be in remote patient monitoring and patient data retrieval. Zhu and Wang [32] present a PDA-based system. It comprises a designed ECG sig-nal acquisition module, a pocket PC, and a GPRS communica-tion module. This system can acquire ECG signals, display the ECG waveform on the liquid crystal display (LCD) screen of the PDA, and wirelessly transmit them to an authorized remote medical ECG management server. Salvador et al. [33] present a platform to enable patients with chronic heart disease to com-plete specifically defined protocols for out-of-hospital follow-up and monitoring. The patients belonged to one of four specific risk groups: arterial hypertension, malignant arrhythmias, heart failure, and postinfarction rehabilitation. They were provided with portable recording equipment and a cellular phone that supported data transmission of ECG and WAP. Fensli et al. [34] present a wireless ECG system for continuous monitoring of ECG activity, especially to diagnose arrhythmia. The patient is required to wear an ECG sensor, which is a smart electronic electrode, with wireless transmission of ECG signals to a dedi-cated handheld device.

All these systems can trace patient’s ECG and display in-formation on the specific devices. However, these systems can still be improved. In case of an emergency, the system sends a warning message, and also reasoning abnormal ECG signal to help medical personnel to diagnose remotely. Beside, since elderly do not necessarily stay in hospitals, we need to identify the elderly person precisely. In this paper, we will overcome the deficiencies and use RFID identification with FPN reasoning in our proposed embedded mobile ECG reasoning system for an elderly patient who lives alone.

Fig. 1. System schematic of a mobile telemedicine system. III. SYSTEMOVERVIEW

The overall system architecture of the proposed telemedicine system is shown in Fig. 1. This telemedicine system combines wireless networks to build the environment for surveillance of critical patients. Our proposed telemedicine services are based on client–server architecture. There is a server (normally located in a hospital) that stores the incoming ECG signals from the clients and makes them available. The client, at an elderly or a patient’s house, is responsible for acquiring data from the patient monitor or the RFID reader and transmitting it through the Internet.

At the patient’s side, the elderly is connected to the patient monitor, which is used to acquire ECG signals. The RFID sys-tem contains a tag with an encoded form of the elderly patient’s identification number, and is tied around the elderly patient’s wrist. The RFID reader converts the elderly patient’s informa-tion from a radio wave into a digital form. The output of elderly information is driven into an embedded system through the RS232 interface. The embedded system communicates between the medical sensor network and the mobile GPRS interface.

This mobile GPRS interface forwards the signals to the telemedicine service provider through GPRS/universal mobile telecommunications system network (UMTS). The server lo-cated in hospital stores data in a relational database. Then, health care providers monitor their patients by the application of the server. The transferred data format of our proposed system is shown in Fig. 2.

In addition, the overall telemedicine services, as shown in Fig. 3, consist of the following four units.

A. Patient Unit

The patient unit consists of the following modules.

1) Module for acquiring elderly patient’s ECG signals: Most of the existing patient monitors are able to acquire tradi-tional and reliable signals. In our prototype, the core mod-ule is an 8-bit microcontroller, PIC16F877, which has an on-chip, eight-channel, 10-bit, analog-to-digital converter (ADC). The three-lead ECG signals were amplified with a gain of 700, filtered (0.5–50 Hz), and fed into the

in-puts of the ADC in the microcontroller. In order to reduce the computation cost of the server, the ECG signals are processed at the client’s end. After processing, only eight important extracted features will be sent to the manage-ment unit for reasoning.

2) Module for elderly patient’s identification: Wristband RFID tag has been used to uniquely identify an elderly. Ev-ery elderly is assumed to have a wristband RFID tag tied around his/her wrist. For example, a PDC Smart Wrist Tag [35] is proposed in this prototype. Each RFID tag contains a unique number that identifies a specific item, called electronic product code (EPC); we use the EPC code to identify the elderly [36]. This tag stores the el-derly patient’s unique identification number. The RFID reader, DR1000 Dual Reader [35], reads the radio signal from the wrist tag. This wristband RFID tag’s data are used to extract the elderly patient’s information from the remote server.

3) Module for data gathering: The embedded system is used to communicate between the medical sensor network and hospital information system (HIS) interface. The input data of elderly patients, manipulated from the RFID reader and the patient monitor, are sent to the embedded sys-tem and transmitted to the server, as shown in Fig. 2. A WISCORE Enterprises’ NET-Start! IXP embedded sys-tem [33] is chosen for this prototyping, which are shown in the lower part of Fig. 4.

4) Module for Internet access: The client application, using a EUPFIN Technology’s EGD-01 GPRS module [35] and running over the Internet through GPRS module, is shown in the upper part of Fig. 4. It provides the core application functionality required by mobile applications.

B. Hospital Unit

The HIS consists of the hospital server and the hospi-tal database system. The hospihospi-tal unit is independent of the telemedicine project in almost all the hospitals. It already exists in the hospital or the private clinic that each pilot cooperates with.

Fig. 2. Transferred data format of our proposed system.

Fig. 3. Architecture of the proposed telemedicine system.

C. Clinician Unit

The clinical unit provides the clinicians with the means to use the telemedicine services. The clinician terminal can receive el-derly patients’ ECG data from the remote patient units via the GPRS/UMTS or the Internet. The clinical unit can provide a continuous monitoring of the ECG of elderly patients. Parameter values and monitor text messages, including alarm conditions, can also be displayed. Moreover, we also consider the confirma-tion mechanism in this unit. It ensures that someone will deal with the emergency event, and the emergency event will not be ignored. In order to intuitionally reflect the health condition, our system uses three colors to represent current condition of elderly. Clinicians can view the status of the elderly patient in-stead of reading the ECG signals. For example, if the elderly

Fig. 4. Embedded system and GPRS module.

Fig. 5. Main control window of the management unit.

patient’s ECG signal is abnormal, then a red signal is shown. Our system provides a PDA-based and smart-phone-based user interface. The upper area represents current condition of an el-derly patient. Patients’ profile is displayed in the middle area. Patients’ ECG signal data are displayed in the lower area.

D. Management Unit

The management unit mainly consists of a fixed PC server and a management program. It is normally located in the nurse’s or clinician’s station, and provides a user-friendly interface for telemonitoring an elderly patient’s ECG signals. The manage-ment program, as shown in Fig. 5, receives the data from the

mobile unit, displays them on the screen, and stores them in a local database. Both the patient unit and the management unit have an alarm setting window, which enables the medical staff to set up an alarm according to the ECG status of the elderly.

In the management unit, the front-end integration service links to the clinician’s Web interface. Also, a back-end integra-tion service links to HIS, which manipulates medical records and other data/knowledge repositories. The profile module contains elderly patients’ medical records and their requirements of the alert module. All elderly patients’ ECG signals are encoded for the input of FPN. Reasoning detail of a patient’s condition is shown in the next section.

IV. ECG SIGNALREASONING

A correct and intuitive representation of the relationship be-tween heartbeats not only makes diagnosis easier, but also helps in understanding the underlying correlation between ECG sig-nals and the symptoms of heart diseases. Therefore, the FPN is used to identify different kinds of abnormal heartbeat. A rea-soning example is illustrated in this section.

A. Dataset Description

PhysioNet [37] provides a set of databases that group records of one or more digitized ECG signals, as well as a set of their cor-responding beat annotations. We used Massachusetts Institute of Technology (MIT)–Beth Israel Hospital (BIH) Arrhythmia Database [38] as the subdatabase to study different types of heartbeats.

The MIT-BIH Arrhythmia Database contains 48 half-hour excerpts of two-channel ambulatory ECG recordings, obtained from 47 subjects studied by the BIH Arrhythmia Laboratory between 1975 and 1979. Twenty-three recordings were chosen at random from a set of four-thousand 24-h ambulatory ECG recordings collected from a mixed population of inpatients and outpatients at Boston’s Beth Israel Hospital. The remaining 25 recordings were selected from the same set that had uncommon but clinically significant arrhythmias, which could not be well represented in a small random sample. These groups comprise men and women between the age of 23 and 89, and are analyzed by two independent cardiologists who classify them on the basis of the types of beats and rhythms. Our selected heartbeat clusters are shown in Table II.

1) ECG Filtering and Sampling: First, we use a nonlinear

bilateral filter to remove noise. It can smooth the noise, but preserve the important features. In image-processing domain, some researchers use this filter to smooth images while preserv-ing edges by means of a nonlinear combination of nearby image values [56]. A bilateral filter is given by

ˆ X(k) =
w n =
−wW (k, n)X(k− n) w n =−wW (k, n) (1) where ˆX(k) is the original signal, ˆX(k) is the smoothed signal,

and w is the width of the filter W(k, n). W (k, n) is the kernel func-tion of the bilateral filter, which is the product of two Gaussion smooth functions and can be expressed as

TABLE II

TYPE OFHEARTBEAT ANDITSDESCRIPTION

TABLE III

TENTIMESNONCONSTANTSAMPLING

W (k, n) = WS(k, n)WR(k, n) (2) where WS(k, n) = exp  − n2 2σ2 S  (3) is the Gaussion kernel in the spatial domain, and

WR(k, n) = exp  −[X(k)− X(k − n)]2 2σ2 R  (4) is the Gaussion kernel in the intensity domain.

The sampling size of our dataset is been estimated by the population ratio [54]

n =z

2

α /2P (1− P )

e2

where n is the sampling size, P is the estimator, e is the sampling error, α is the confidence coefficient, and z is the standard normal probability distribution. Since there is no further study on the estimation of an ECG, beat will represent the abnormal beat’s probability (P will never be known exactly). Therefore, we use nonconstant number sampling to obtain the abnormal beat’s proportion of population in a ten-time trial, which is shown in Table III. From Table III, we can obtain an abnormal beat’s estimator (population proportion expected value)

P =

10



i= 1

TABLE IV

SAMPLINGNUMBER INDIFFERENTCLUSTERS

TABLE V

FEATURES OFHEARTBEAT ANDITSDESCRIPTION

Substituting for the value of P, with 0.95 confidence levels and 0.02 sampling error, the sampling beat’s size n is estimated to be 553. By sampling beat in different groups, as shown in Table IV, we obtained 553 beats (120 normal and 433 abnormal) in our sampling dataset. About two-thirds (67.5%) of the dataset are training data that are used to train the knowledge model and one-third (32.5%) of the dataset is used to evaluate the model.

2) Feature Extraction: The selection of suitable features is

important in the automatic classification of ECG. Some re-searchers use different features in frequency domain, and others calculate the features in the time domain. From our experiences, we found that the features listed in Table V can better perform in our prototyping. Nevertheless, moving average method and the differential equation approach are used in our paper. The steps involved in detecting peaks are: 1) calculating the mov-ing average from original signal, and usually usmov-ing the past ten records to calculate it; 2) subtracting moving average from orig-inal signal, and obtaining a new signal; 3) finding the peak of the signal; and 4) setting the threshold to decide the detected peaks. As soon as the peak “R” is obtained, peaks P, Q, S, and T can also be found through their relative positions. In order to prevent possible error, the peaks’ amplitude is measured from k line, as shown in Fig. 6. The definition of a k line is given as

k = Max(θi, i = 1, 2, . . . , 11) + c (5)

where k is a baseline, θ is the greatest amplitude of all peaks,

i is the type of heartbeat, and c is a constant. Therefore, the

amplitude of peaks P, Q, R, S, and T are all homogeneously negative.

Fig. 6. Single ECG wave with k baseline.

Fig. 7. Definition of FPN.

B. Fuzzy Petri Net Model

FPN can be used to describe most of the relations and be-haviors in a discrete-event dynamic system. As a mathematical tool, behavior properties and performance evaluation can be conveniently analyzed by using net theory and algebraic theory. Therefore, the reasoning path of expert systems can be reduced to simple sprouting trees if FPN-based reasoning algorithms are applied as an inference engine. Many results prove that FPN is suitable for representing and reasoning misty logic implication relations [30]. A generalized FPN structure can be defined as an eight-tuple in Fig. 7 [39], [40].

Fuzzy production rules (FPRs) are usually presented in the form of a fuzzy IF–THEN rule in which both the antecedent and the consequent are fuzzy concepts denoted by fuzzy sets. To effectively represent both the fuzziness and the uncertainty in FPRs, several knowledge parameters such as the values of certainty factor (CF) (µi) and threshold (λ) have been

incorpo-rated into the FPRs [39], [41], [42]. The composite FPR can be distinguished into the following rule types [39].

Type 1: IF dj THEN dk (CF = µi). The fuzzy reasoning

pro-cess of this type of rule can be represented by

Fig. 8. Simple FPR.

Type 2: IF dj 1and dj 2 and· · · and dj n THEN dk (CF = µi).

The fuzzy reasoning process of this type of rule can be represented by

α (Pk) = min{α (Pj 1) , α (Pj 2) , . . . , α (Pj n)} ∗ µi. (7)

Type 3: IF dj THEN dk 1and dk 2and· · · and dk n (CF = µi).

The fuzzy reasoning process of this type of rule can be represented by

α (Pk 1) = α (Pj)∗ µi, α (Pk 2)

= α (Pj)∗ µi, . . . , α (Pk n) = α (Pj)∗ µi. (8)

Type 4: IF dj 1 or dj 2 or · · · or dj n THEN dk (CF = µi).

The fuzzy reasoning process of this type of rule can be represented by

α (Pk) = max{α (Pj 1)∗µi, α (Pj 2)∗µi, . . . , α (Pj n)∗µi} .

(9) Type 5: IF dj THEN dk 1 or dk 2 or · · · or dk n (CF = µi).

The fuzzy reasoning process of this type of rule can be represented by

α (Pk 1)=α (Pj)∗ µi1, α (Pk 2) =α (Pj)∗ µi2, . . . , α (Pk n)

= α (Pj)∗ µin. (10)

A simple FPR is mathematically and graphically illustrated in Fig. 8. Given a set of FPRs and an observed fact, FPR reasoning is used to draw an approximate conclusion by matching the observed fact against the set of FPRs [39], [42]. The truth of each proposition is repressed by fuzzy membership functions and the CFs for the propagation of uncertain evidence from the antecedent of a production rule to its consequences [41], [55].

To subdivide our dataset into membership functions, we de-termine a threshold line with entropy-minimization screening method [43], and then start the segmentation process. We can gain the degree of truth of each feature under some states through the fuzzy membership function [40]. Each of the features is par-titioned into three clusters: high, median, and low. For exam-ple, the value of the RH is in the range between−2.1284 and

−2.8311, and this value meets two membership functions L1

and L2. Then, the maximum membership degree is chosen for

RH. Fig. 9 is the fuzzy membership function of the RH, and the fuzzy membership functions of other features are obtained similarly.

In developing the CF model, Buchanan and Shortliffe [41] chose two basic measures of uncertainty: the measure of belief MB[h, e] that expresses the degree to which an observed piece of evidence e increases the belief in a hypothesis h, and the measure of disbelief MD[h, e] that expresses the degree to which an observed piece of evidence e decreases the belief in a hypothesis

h. Each of these measures lies in the closed interval [0, 1].

The measure of belief MB[h, e] and the measure of disbelief MD[h, e] are defined as MB[h, e] =    1, if P (h) = 1 max [P (h| e), P (h)] − P (h) max [1, 0]− P (h) , otherwise (11) MD[h, e] =    1, if P (h) = 0 min [P (h| e), P (h)] − P (h) min [1, 0]− P (h) , otherwise (12) where P (h| e) and P(h) represent conditional and a priori prob-abilities, respectively.

In addition, a CF is definite, as shown in the following equation:

CF[h, e] = MB[h, e]− MD[h, e] (13) and the function for combining two CFs CF(e1, e) and

CF(e2, e) of two constituting pieces of evidence e1is

CF(e1 and e2, e) = min{CF(e1, e), CF(e2, e)}. (14)

For the disjunction of two pieces of evidence, we have CF(e1 or e2, e) = max{CF(e1, e), CF(e2, e)} (15)

and the combination function for combining two CFs, CF(h, e₁) and CF(h, e₂), which have been derived from two coconcluding production rules, is shown in the following:

CF(h, e₁co e₂) =        x + y(1− x), if x, y > 0 x + y 1− min {|x| , |y|}, if − 1 < x, y ≤ 0 x + y(1 + x), if x, y < 0 (16) where CF(h, e1) = x and CF(h, e2) = y.

By combining the propositions and CFs, we can obtain a set of FPRs, and the structure of our FPN for the type of heartbeats reasoning is shown in Fig. 10.

C. FPN Reasoning

From Fig. 10, our FPN for heartbeats reasoning have eight inputs (eight features), 11 middle states (1 normal heart beat and 10 abnormal heartbeats), and one final state. All the degree of truth and states of the propositions in the FPN are gained by the fuzzy membership functions. All the values of CF in the FPN are calculated by (11)–(13). When the CF is used, the propaga-tion of uncertainty in the inference network is accomplished by repeated application of sequential and parallel combination of CFs. According to the different antecedent of a production rule,

Fig. 9. Fuzzy membership function of RH.

Fig. 10. FPN reasoning model.

different combination functions are used for the propagation of uncertain evidence. The knowledge base of our proposed system contains the following FPRs, R1–R12, as shown at the bottom of this page.

All the ECG samples acquire membership degree and linguistic variable themselves from the input of FPN by the fuzzy mem-bership function. We can obtain the possible cluster and degree of an ECG signal by reasoning process.

For instance, given an ECG signal, the membership function returns the membership degree of eight characteristics as [0.84, 0.54, 0.73, 0.56, 1.00, 0.91, 0.63, 0.85], and the fuzzy linguistic variable of [x1, x2, . . ., x8] are [M, M, M, M, M, L, H, M]. Let

β(P1) = d1and β(P2) = d2, . . ., β(P8) = d8, where the places

P1,. . ., P8are called the starting place and d1, d2, d3, . . ., d8are

eight propositions of ECG signals. The variables d9, d10, . . .,

d19denote 11 propositions of ECG clusters, and the goal place

and propositions are P20 and d20, respectively. The reasoning

example of FPN is as follows.

At first, the fuzzy linguistic variables and the degree of truth of place P1, P2, . . ., P8 can be gained by fuzzy membership

function.

Step 1: Starting places are P1, P2, and P8: tokens are moved

from P1, P2,. . ., P8 to P9, P10,. . ., P19

α(Pj) = Min{α(Pi)∗ tij|i = 1, 2, . . . , 8}, where

j = 9, 10, . . . , 19.

Therefore, the proposition of place P9, P10,. . ., P19

are

{d9, d10, . . . , d19}

={0.2532, 0.1674, 0.2646, 0.1953, 0.3136, 0.2352, 0.2961, 0.3584, 0.448, 0.42, 0.3456}.

Step 2: Tokens are moved from P9, P10,. . .,P19 to P20.

As-sume that all of the CFs linked to P20are 0.8, ti,20=

0.8, and the proposition of place P20 is d20(k), the

truth of d20(k) can be gained by FPR as

d20(k) ={α(Pi)∗ ti,20|i = 9, 12, . . . , 19}, k = 1, 2, . . . , 11 ={α(P9)∗ t9,20, α(P10)∗ t10,20, α(P11)∗ t11,20, α(P12)∗ t12,20, α(P13)∗ t13,20, α(P14)∗ t14,20, α(P15)∗ t15,20, α(P16)∗ t16,20, α(P17)∗ t17,20, α(P18)∗ t18,20, α(P19)∗ t19,20}.

Step 3: Since there are 11 success paths in final place P20,

we can select the degree of truth of place P20=

Max{d20(k)} and the proposition β(P20) by FPRs

R1 : IF d1 and d2· · · and d8 THEN d9 R7 : IF d1 and d2· · · and d8 THEN d15

R2 : IF d1 and d2· · · and d8 THEN d10 R8 : IF d1 and d2· · · and d8 THEN d16

R3 : IF d1 and d2· · · and d8 THEN d11 R9 : IF d1 and d2· · · and d8 THEN d17

R4 : IF d1 and d2· · · and d8 THEN d12 R10 : IF d1 and d2· · · and d8 THEN d18

R5 : IF d1 and d2· · · and d8 THEN d13 R11 : IF d1 and d2· · · and d8 THEN d19

TABLE VI

CONFUSIONMATRIX OFOURBEATCLASSIFICATION

as

d20 = Max{d20(k)}, k = 1, 2, . . . , 11

= Max{0.2026, 0.1339, 0.2117, 0.1562, 0.2509, 0.1882, 0.2369, 0.2867, 0.3584, 0.3360, 0.2765} = 0.3584 (at P17 → P20).

Thus, β(P20) = cluster 9, and this ECG signal belongs

to cluster 9 in Table II.

Therefore, we can obtain the cluster to which the ECG signal belongs to, and degree of high-risk heartbeats is 0.3584 from this example. If this degree exceeds a predefined range, the precon-figured red, yellow, or green light will be activated accordingly, and doctors will receive warning message and will immediately take care of the elderly.

D. Evaluation of FPN Reasoning

In order to evaluate the performance of our proposed method, the measuring criterion is defined as follows:

True positives (TP) number of correct beats classified as correct cluster;

True negatives (TN) number of incorrect beats classified as incorrect cluster;

False positives (FP) number of incorrect beats classified as correct cluster;

False negatives (FN) number of correct beats classified as incorrect cluster.

And, the overall accuracy is defined with equation (TP +TN)/(TP +FP + FN + TN).

In our evaluation, we use 368 cases to train our proposed FPN model and 185 cases for testing from sampled dataset of Table IV. Table VI summarizes the results of the confusion matrix that displays classification of beat recognition using our proposed FPN reasoning model. Table VII summarizes the com-parative results of the accuracy of beat recognition with support vector machine classifier [23], neuro-fuzzy network [24], fuzzy hybrid neural network [18], and mirrored Gauss model [22]. We can successfully recognize most of clusters since every cluster’s

TABLE VII

COMPARATIVERESULTS OFBEATRECOGNITION

discriminate accuracy is greater than 90%. The comparison in-dicates good accuracy of our proposed method.

V. SYSTEMEVALUATION ANDDISCUSSION

When organizations decide to adopt a new technology or an information system, organizations have to address acceptance issues such as resistance to change, compatibility, and fear of adverse consequences [45], [46]. Social psychology literature includes the technology acceptance model (TAM) and the the-ory of planned behavior (TPB), which address the various prob-lems that arise when a new information system or technological innovation in the information technology (IT) field is introduced in the workplace. These models have received significant atten-tion in the research literature. The TAM, especially, has been applied in many usage scenarios and has performed consistently well. Based on social psychology, TAM is a parsimonious and powerful model for predicting the usage of an information sys-tem. TAM claims that ease of use and usefulness can predict usage of a system. Attitude and behavioral intention act as the mediating variables in TAM [47]. The TPB is a particularly well-established intention model that has been proved success-ful in predicting and explaining behavior across a wide variety of domains [48], [49]. The TPB posits attitude as an antecedent of intention to perform behavior. Previous research has found that when constructs of these or similar models are combined, they show better predictability compared to that of the each model alone [47], [50].

In order to understand the acceptance of our system and the drawbacks of the system, we propose a new research model, which modified constructs from the TAM and TPB, and in-cludes “mobility,” “usability,” “performance of the overall

sys-tem,” “attitude,” and “intention to use” in this context. In our

research model, three important factors that impact individual’s attitude are “mobility,” “usability,” and “performance of the

overall system.” Attitude toward using this system is the user’s

evaluation of the desirability of his\her using the system and the individual’s positive or negative feelings about performing the target behavior. Based upon the empirical research of TAM constructs, the proposed model studies the impacts of these con-structs on the consumer adoption patterns. Our research model is shown in Fig. 11. The hypotheses for the test are as follows.

Fig. 11. Research model.

H1: The mobility of the system and the attitude of users are

positively correlated.

H2: The usability of the system and the attitude of users are

positively correlated.

H3: The performance of the overall system and the attitude

of users are positively correlated.

H4: The attitude of users and intention to use are positively

correlated.

The prototyping system was demonstrated at our laboratory. In the test scenario, 25 volunteers were asked to use our sys-tem, and eight interns stayed in a room in another building and observed the ECG signal of the elderly patient. The patient unit transmits the elderly patient’s ECG signals and patient’s ID to the management unit via GPRS, thus allowing medical staffs to monitor online.

According to the test scenario, a survey was conducted to elicit the operators’ opinions on the mobile classification of the ECG reasoning system in five areas, which are mobility, usabil-ity, performance of the overall system, attitude of users, and intention to use. A questionnaire with a five-point Likert scale (from 5 = completely satisfied to 1 = completely unsatisfied) was used to rate the performance of the overall system. The satisfaction of mobility was evaluated in relation to statements of weight and size, and easy operation and easy monitoring es-timated usability. The attitude of users and intention to use were also assessed.

Since the IT professionals and doctors are the major adopters in the medical organization, 25 volunteers and eight interns, in-cluding 22 males and 11 females, took part in the experiment and answered the questionnaire. All volunteers were undergrad-uates with the age ranging between 19 and 22 years. The age of the intent was within the range of 28–35 years (mean = 31.4, SD = 2.6).

The SPSS package was used to compute frequencies, means, percentage, reliability test, and Pearson correlation analysis. The Cronbach’s coefficient was used to determine the reliability of questionnaire items. Table VIII shows the values of alpha. Since the values of Cronbach’s coefficient are greater than 0.70, the recommended value is based on [44], and the reliability of construct for the sample in this study was considered to be sufficient.

We tested whether there was a positive correlation between mobility of the system, usability of the system, and performance of the overall system toward the attitude of users. We also

ex-TABLE VIII

RELIABILITYANALYSIS(CRONBACH’SALPHA)

TABLE IX

CORRELATIONSBETWEENMAINCONSTRUCTS

amined whether there was a positive correlation between the attitude of users and personal intent to use the system. The hy-potheses (H1− H4) were tested using Pearson correlation

pro-cedure. The results of the correlations and their significances are shown in Table IX.

Based on the Pearson test, all four hypotheses (H1− H4)

were accepted, i.e., the user’s attitude was significantly corre-lated with user’s intent to use the system, and all factors

(“mobil-ity,” “usabil(“mobil-ity,” and “performance of the overall system”) were

significant in affecting user’s attitude toward using the system. The “usability” was the strongest variable that affected attitude toward system use, followed by “performance of the overall

system” and “mobility.”

Factor analysis can also be performed to confirm that the items are loaded according to the proposed model. According to [51], the acceptable value for factor loading is greater than 0.5. Thus, items with loading less than 50% were dropped from further analysis. Appendix A presents the results of factor anal-ysis. Even more, we can employ a structural equation modeling (SEM) technique to analyze the research model we proposed. However, these data analyses are beyond our topic. They need further investigation and are not included in this paper. We will proceed with all the possible research in the near future.

Volunteers have lower mean values than doctors’ in attitude and intention to use constructs. Roughly speaking, those interns highly rated the overall system for performance on our mobile classification of the ECG reasoning system. This study pro-vided us an opportunity to express user’s thoughts and feelings that they had experienced while using our prototyping system. Some of the suggestions for further improvement included that we make our system more complete, such as security and pri-vacy consideration, the patient unit only periodically reports to the management unit, etc. In order to increase security and pri-vacy in the future, we plan to incorporate encryption technology into our system such as hash function, asymmetric and sym-metric encryption/decryption, key shadow, etc. Another major drawback was one-way communication between the embedded system and the HIS server. Only embedded system can send

TABLE X

COMPARATIVERESULTS OFOTHERSYSTEMS

data to server; the server cannot send any message to control the embedded system. In the future, we will use the SMS ser-vice to interact with embedded system and server. For exam-ple, the command format is (#Time, ID, PassWd, Command, [Parameter∗] #). All the instructions are enclosed within two “%” signs. The ID and PassWD stand for server ID and pass-word; the instruction issued by Command and Parameter will be executed. Then, full duplex communication between the em-bedded system and the HIS server can be achieved.

VI. CONCLUSION

A monitoring system that performs a complete ECG analysis in a local device near the elderly patients is of great interest since it allows us to improve the quality of life of cardiovas-cular patients. By applying wristband RFID tag on an elderly, the medical data can be permanently attached to each elderly patient at all times. This art of composition greatly reduces the possibility of errors, and hence helps speed up the treatment in the hospital or at home. Our proposed system can allow much quicker treatment during emergency even before the elderly pa-tient arrives at the hospital. Moreover, our FPN reasoning model can have accuracy rate on beat type from 90.8% to 97.8%. If system detects an abnormal ECG activity, related hospitals will receive the alarm message even more precisely. The comparison with other systems’ environment is also shown in Table X. In addition, we surveyed 33 people to test whether there was a positive correlation between the attitude of users and personal intent to use the system. The results support that mobility, us-ability, and performance of our proposed system have impacts on user’s attitude, and there is a significant positive effect be-tween the user’s attitude and the intent to use in our proposed system.

Mobile telemedicine is an emerging area that has the potential to change the way of health care provided today [52]. Although our proposed FPN model shows a good accuracy in beat recog-nition, however, it still suffers from self-learning capability and state explosion problem. In the future, we will try to investigate encryption technology and provide a collaborative framework that incorporates several different classification methods to im-prove our system’s drawbacks, and further, it may incorporate EEG vital signal reasoning. Therefore, the system would pro-vide a better quality of care for the elderly patient who lives alone. Another important research issue with health monitoring is the privacy and security aspect. Experts from Berkeley be-lieve that security must be integrated into every component, as components designed without security are the vulnerable points where attacks start [53].

APPENDIXA

ROTATEDCOMPONENTMATRIX

REFERENCES

[1] D. Bottazzi, A. Corradi, and R. Montanari, “Context-aware middleware solutions for anytime and anywhere emergency assistance to elderly peo-ple,” IEEE Commun. Mag., vol. 44, no. 4, pp. 82–90, Apr. 2006. [2] C. C. Lin, M. J. Chiu, C. C. Hsiao, R. G. Lee, and Y. S. Tsai, “A wireless

healthcare service system for elderly with Dementia,” IEEE Trans. Inf.

Technol. Biomed., vol. 10, no. 4, pp. 696–704, 2006.

[3] Irishhealthhealth. (2006). [Online]. Available: http://www.irishhealth. com/index.html?level=4&id=2013

[4] F. Lamberti and C. Demartini, “Low-cost home monitoring using a Java-based embedded computer,” in Proc. 4th Int. IEEE EMBS Spec. Topic,

Inf. Technol. Appl. Biomed., 2003, pp. 342–345.

[5] J. Garcia, I. Martinez, L. Sornmo, S. Olmos, A. Mur, and P. Laguna, “Remote processing server for ECG-based clinical diagnosis support,”

IEEE Trans. Inf. Technol. Biomed., vol. 6, no. 4, pp. 277–284, Dec. 2002.

[6] M. Al-Rousan, A. R. Al-Ali, and A. Eberlein, “Remote patient monitoring and information systems,” Int. J. Electron. Healthcare, vol. 2, no. 3, pp. 231–249, 2006.

[7] S. Sneha and U. Varshney, “A wireless ECG monitoring system for per-vasive healthcare,” Int. J. Electron. Healthcare, vol. 3, no. 1, pp. 32–50, 2007.

[8] B. Wu, Z. Yu, X. Zhu, Q. Yan, L. Zhu, and G. Li, “A novel mobile ECG telemonitoring,” in Proc. 27th Annu. Int. Conf. IEEE-EMBS, Sep.1–4, 2005, pp. 3818–3821.

[9] J. Muhlsteff, O. Such, R. Schmidt, M. Perkuhn, H. Reiter, J. Lauter, J. Thijs, G. Musch, and M. Harris, “Wearable approach for continuous ECG–and activity patient-monitoring,” in Proc. 26th Annu. Int. Conf.

EMBC, 2004, vol. 1, pp. 2184–2187.

[10] Prognosis Good for Wireless in Health Care. (2005). [Online]. Available: http://www.thefeaturearchives.com/100929&ref=7285844.html [11] W. D. Yu and A. Ramani, “Design and implementation of a personal

mo-bile medical assistant,” in Proc. Int. Workshop Enterprise Netw. Comput.

Healthcare Ind., 2005, pp. 172–178.

[12] P. E. McSharry, G. D. Clifford, L. Terassenko, and L. A. Smith, “A dy-namical model for generating synthetic electrocardiogram signals,” IEEE

Trans. Biomed. Eng., vol. 50, no. 3, pp. 289–294, Mar. 2003.

[13] B. N. Singh and A. K. Tiwari, “Optimal selection of wavelet basis function applied to ECG signal denoising,” Dig. Signal Process., vol. 16, no. 3, pp. 275–287, May 2006.

[14] C. Philip de, M. O’Dwyer, and R. B. Reilly, “Automatic classification of heartbeats using ECG morphology and heartbeat interval features,” IEEE

Trans. Biomed. Eng., vol. 51, no. 7, pp. 1196–1206, Jul. 2004.

[15] T. P. Exarchos, C. Papaloukas, D. I. Fotiadis, and L. K. Michalis, “An association rule mining based methodology for automated detection of Ischemic ECG beats,” IEEE Trans. Biomed. Eng., vol. 53, no. 8, pp. 1531– 1540, Aug. 2006.

[16] H. G. Hosseini, D. Luo, and K. J. Reynolds, “The comparison of different feed forward neural network architectures for ECG signal diagnosis,”

Med. Eng. Phys., vol. 28, no. 4, pp. 372–378, May 2006.

[17] Z. Dokur and T. ¨Olmez, “ECG beat classification by a novel hybrid neural network,” Comput. Methods Programs Biomed., vol. 66, no. 2/3, pp. 167– 181, Sep. 2001.

[18] S. Osowski and T. H. Linh, “ECG beat recognition using fuzzy hybrid neural network,” IEEE Trans. Biomed. Eng., vol. 48, no. 11, pp. 1265– 1271, Nov. 2001.

[19] T. H. Linh, S. Osowski, and M. Stodolski, “On-line heart beat recogni-tion using Hermite polynomials and neuro-fuzzy network,” IEEE Trans.

Instrum. Meas., vol. 52, no. 4, pp. 1224–1231, Aug. 2003.

[20] P. M. Yeong, F. Ibrahim, S. A. L. Narainasamy, and R. Omar, “Intelligent classification of electrocardiogram (ECG) signal using extended Kalman Filter (EKF) based neuro fuzzy system,” Comput. Methods Programs

Biomed., vol. 82, no. 2, pp. 157–168, May 2006.

[21] R. V. Andreao, B. Dorizzi, and J. Boudy, “ECG signal analysis through hidden Markov models,” IEEE Trans. Biomed. Eng., vol. 53, no. 8, pp. 1541–1549, Aug. 2006.

[22] Q. Zhou, X. Liu, and H. Duan, “ECG beat classification using Mirrored Gauss model,” in Proc. 27th Annu. Conf. IEEE Eng. Med. Biol., China, Sep. 1–4, 2005, pp. 5587–5590.

[23] N. Acir, “A support vector machine classifier algorithm based on a per-turbation method and its application to Implementation of a WAP-Based Telemedicine System for Patient Monitoring,” Exp. Syst. Appl., vol. 31, no. 1, pp. 150–158, Jul. 2006.

[24] M. Engin, “ECG beat classification using neuro-fuzzy network,” Pattern

Recognit. Lett., vol. 25, no. 15, pp. 1715–1722, Nov. 2004.

[25] G. Inan and E. D. ¨Ubeyl, “ECG beat classifier designed by combined neural network model,” Pattern Recognit., vol. 38, no. 2, pp. 199–208, Feb. 2005.

[26] K. Polat, S. Sahan, and S. G¨unes, “A new method to medical diagnosis: Artificial immune recognition system (AIRS) with fuzzy weighted pre-processing and application to ECG arrhythmia,” Exp. Syst. Appl., vol. 31, no. 2, pp. 264–269, Aug. 2006.

[27] C. Lin, C. M. Lin, B. Lin, and M. C. Yang, “A decision support system for improving doctors’ prescribing behavior,” Exp. Syst. Appl., vol. 38, no. 4, pp. 7975–7984, May 2009.

[28] L. Zhang, L. Wang, Y. Zang, H. Seki, and H. Itoh, “On rule checking and learning in an acupuncture diagnosis fuzzy expert system by genetic algorithm,” in Proc. IEEE Int. Conf. Fuzzy Syst., Yokohama, Japan, 1995, vol. 2, pp. 455–460.

[29] P. A. Heng, T. T. Wong, R. Yang, Y. P. Chui, Y. M. Xie, K. S. Leung, and P. C. Leung, “Intelligent inferencing and haptic simulation for Chinese acupuncture learning and training,” IEEE Trans. Inf. Technol. Biomed., vol. 10, no. 1, pp. 28–41, Jan. 2006.

[30] X. Li and W. Yu, “Dynamic knowledge inference and learning under adaptive fuzzy Petri net framework,” IEEE Trans. Syst., Man, Cybern. C,

Appl. Rev., vol. 30, no. 4, pp. 442–450, Nov. 2000.

[31] K. Hung and Y.-T. Zhang, “Implementation of a WAP-based telemedicine system for patient monitoring,” IEEE Trans. Inf. Technol. Biomed., vol. 7, no. 2, pp. 101–107, Jun. 2003.

[32] Q. Zhu and M. Wang, “A wireless PDA-based electrocardiogram trans-mission system for telemedicine,” in Proc. 27th Annu. Conf. IEEE Eng.

Med. Biol., Sep.1–4, 2005, pp. 3807–3809.

[33] C. H. Salvador, M. P. Carrasco, M. A. G. de Mingo, A. M. Carrero, J. M. Montes, L. S. Martin, M. A. Cavero, I. F. Lozano, and J. L. Monteagudo, “Airmed-cardio: A GSM and Internet services-based system for out-of-hospital follow-up of cardiac patients,” IEEE Trans. Inf. Technol. Biomed., vol. 9, no. 1, pp. 73–85, Mar. 2005.

[34] R. Fensli, E. Gunnarson, and O. Hejlesen, “A wireless ECG system for continuous event recording and communication to a Clinical Alarm Sta-tion,” in Proc. 26th Annu. Int. Conf. EMBC, 2004, vol. 1, pp. 2208–2211. [35] EUPFIN Technology. (2006). [Online]. Available: http://www.eup.

com.tw/a5b2.htm

[36] E. Lubrin, E. Lawrence, and K. F. Navarro, “Wireless remote healthcare monitoring with Motes,” in Proc. Int. Conf. Mobile Bus., 2005, pp. 235– 241.

[37] PhysioBank. (2006). [Online]. Available: http://www.physionet.org/ physiobank/

[38] PhysioBank. (2006). [Online]. Available: http://www.physionet.org/ physiobank/database/mitdb/

[39] S. M. Chen, J. S. Ke, and J. F. Chang, “Knowledge representation using fuzzy Petri nets,” IEEE Trans. Knowl. Data Eng., vol. 2, no. 3, pp. 311– 319, Sep. 1990.

[40] T. J. Ross, Fuzzy Logic With Engineering Applications. New York: McGraw-Hill, 2000.

[41] B. G. Buchanan and E. H. Shortliffe, Eds., Rule-Based Expert Systems:

The MYCIN Experiments of the Stanford Heuristic Programming Project.

Reading, MA: Addison-Wesley, 1984.

[42] D. S. Yeung and E. C. C. Tsang, “Weighted fuzzy production rules,” Fuzzy

Sets Syst, vol. 88, pp. 299–313, 1997.

[43] R. Christensen, Entropy Minimax Sourcebook, Lincoln, MA: Entropy Ltd., 1980.

[44] J. Nunnally, Psychometric Theory. New York: McGraw-Hill, 1978.

[45] C. Flavian and M. Guinaliu, “Consumer trust, perceived secure and privacy policy,” Ind. Manage. Data Syst., vol. 106, no. 5, pp. 601–620, 2006. [46] V. J. Garcis-Morales, F. J. Llorens-Montes, and A. J. Verdu-Jover,

“An-tecedents and consequences of organizational innovation and organiza-tional learning in entrepreneurship,” Ind. Manage. Data Syst., vol. 106, no. 1, pp. 21–42, 2006.

[47] F. Davis, “Perceived usefulness, perceived ease of use, and end user accep-tance of information technology,” MIS Quart., vol. 13, no. 3, pp. 318–339, 1989.

[48] M. K. Chang, “Predicting unethical behavior: A comparison of the theory of reasoned action and the theory of planned behavior,” J. Bus. Ethics, vol. 17, pp. 1825–1834, 1998.

[49] R. A. Johnson, B. C. Hardgrave, and E. R Doke, “An industry anal-ysis of developer beliefs about object-oriented systems development,”

DATABASE, vol. 30, no. 1, pp. 47–64, 1999.

[50] B. Hardgrave and R. Johnson, “Toward an information systems devel-opment acceptance model: The case of object-oriented systems develop-ment,” IEEE Trans. Eng. Manage., vol. 50, no. 3, pp. 322–336, Aug. 2003.

[51] J. Hair, R. Anderson, R. Tatham, and W. Black, Multivariate Data

Anal-ysis. Englewood Cliffs, NJ: Prentice-Hall, 1998.

[52] H. Y. Kung, C. Y. Hsu, M. H. Lin, and C. N. Liu, “Mobile multimedia medical system: Design and implementation,” Int. J. Mobile Commun., vol. 4, no. 5, pp. 595–620, 2006.

[53] D. Culler and W. Hong, “Wireless sensor networks,” Commun. ACM, vol. 47, no. 6, pp. 41–46, 2004.

[54] W. Mendenhall and R. Beaver, Introduction to Probability and Statistics. Belmont, CA: Duxbury Press, 1994.

[55] E. H. Shortliffe, Computer-Based Medical Consultations: MYCIN. New York: Elsevier, 1976.

[56] M. Elad, “On the origin of the bilateral filter and ways to improve it,”

IEEE Trans. Image Process., vol. 11, no. 10, pp. 1141–1151, Oct. 2002.

Dong-Her Shih received the Ph.D. degree in

electri-cal engineering from the National Cheng Kung Uni-versity, Tainan, Taiwan, in 1986.

He is currently a Senior Professor in the De-partment of Information Management, National Yun-lin University of Science and Technology, Douliu, Taiwan. His current research interests include intru-sion prevention system (IPS), machine learning, data mining, security, ontology, and wireless network.

Hsiu-Sen Chiang received the Ph.D. degree in

infor-mation management from the National Yunlin Uni-versity of Science and Technology, Douliu, Taiwan, in 2007.

He is currently an Assistant Professor in the De-partment of Information Management, Da-Yeh Uni-versity, Changhua, Taiwan. His current research in-terests include data mining, ontology, Petri net, and malicious behavior detection.

Binshan Lin (M’09) received the Ph.D. degree in

quantitative business analysis from Louisiana State University, Baton Rouge, in 1988.

He is currently the BellSouth Corporation Profes-sor at College of Business Administration, Louisiana State University, Shreveport. He has authored or coauthored more than 180 articles in refereed jour-nals. He is currently the Editor-in-Chief of several academic journals.

Shih-Bin Lin received the M.S. degree in information

management from the National Yunlin University of Science and Technology, Douliu, Taiwan, in 2006. He is currently working toward the Ph.D. degree at the Department of Information Management, National Chiao Tung University, Hsinchu, Taiwan.

His current research interests include data mining, embedded system, and information retrievals.