Body Area Networks Multi-Attribute Decision Making Handover Algorithm for Wireless Computer Communications

ContentslistsavailableatScienceDirect

Computer Communications

journalhomepage:www.elsevier.com/locate/comcom

Multi-Attribute Decision Making Handover Algorithm for Wireless Body Area Networks

Hadda Ben Elhadj

^a,∗

, Jocelyne Elias

^b

, Lamia Chaari

^a

, Lotfi Kamoun

^a

aLETI Laboratory, Sfax University, Tunisia

bLIPADE Laboratory, Université Paris Descartes – Sorbonne Paris Cité, Paris 75006, France

a rt i c l e i n f o

Article history:

Received 26 November 2014 Revised 7 November 2015 Accepted 23 January 2016 Available online 12 February 2016 Keywords:

Wireless Body Area Networks Healthcare

Handover Quality of Service

a b s t r a c t

Inthispaper,wetackle theWirelessBody areaNetwork(WBAN) handoverissue whereamobile pa- tienthastoselectatanytimethebestaccesstechnologyaccordingtomultiplecriteria.Weparticularly focusonthedecision schemesand investigatethe Multi-AttributeDecision Making(MADM)methods.

ThefundamentalobjectiveoftheMADMmethodsistodetermineamongafinitesetofalternativesthe optimalone.Therefore,weproposeaMulti-AttributeDecisionMakingHandoverAlgorithm(MADMHA) whichhelpspatient’smobile terminal to dynamicallyselect thebestnetworkby providingaranking orderbetweenthelistofavailablecandidates.Itisaseamlesshandoverapproachthatguaranteescon- tinuousconnectivitywithrespecttotheQoSrequirementsoftheWBANgeneratedtraffictypes,network historyanduserpreference.Simulationresultsprovetheefficiencyofourproposedapproachversusthe ReceivedSignalStrengthIndicator(RSSI)andDataRate(DR)basedhandoverapproaches.Indeed,com- paredtotheselatters,MADMHAsignificantlyreducesthepacketoverheadandthenumberofhandover, whilelimitingthepacketlossratio.

© 2016ElsevierB.V.Allrightsreserved.

1. Introduction

WBANs are an effective means to provide many promising applications in different domains [1–3]. In fact, WBANs are ap- plied in a variety of areas such as healthcare, medicine, patient monitoring, sport and multimedia, to cite a few. In the health- care domain, a WBAN consistsofa set of medicalsensors (ECG:

Electrocardiogram, EEG: Electroencephalogram, etc.) implantedin or on the user’s body, and a coordinator for data transmission, whichcan bea PersonalDigital Assistant(PDA)ora smartphone.

These devices collect, store and process patient’s physiological parameters(heartbeat,bloodpressure,bodytemperature,etc.)and provideubiquitoushealthcareservices.

Indeed, remote healthcare monitoring technology is expected to reduceunnecessary hospitalizationsandshortenlength ofstay when admission is necessary. It improves the level of patients engagement and care, regardless of their location around the globe, andenables more timely intervention fromcaregivers and clinicians through real-time data monitoring andalerts. For clar- ification purposes,let ustake the exampleof apatient, suffering

∗ Corresponding author. Tel.: +216 25369105.

E-mail addresses: [email protected] (H. Ben Elhadj), [email protected] (J. Elias), [email protected] (L. Chaari), [email protected] (L. Kamoun).

fromcoronaryheartdisease,whowantstogohomesafeknowing thatheistakencareof.So,heleavesthehospitalandissurethat amedicalteamwillbepromptlydispatched tohislocation when needed. To achieve this, the remote healthcare ECG application used by the patient must send its ECG pattern and location to thehealthcareprofessionalswheneveran irregularECGpatternis detected. Another interesting example of timely interventions is anapplicationusedfortraumasituation.Inthiscase,thesurgeon maytake adecisionaboutthesurgeryona traumapatient based onthecontinuously received bio-signalsatthe hospitalback end healthcare server, while the on-site trauma team is transporting thepatienttothehospital[4].

However, the effectiveness of these real-life trials of remote healthcare applications can be affected by user mobility, wireless networks availability, required healthcare QoS, network density andthebatterylimitationsofmobile devices.Hence,thereis the needforaseamlesshandoverapproachthatensurespatientmobil- itymanagementwhilekeepingthepatientalwaysbestconnected.

Furthermore,since recentPDAandsmartphones are equipped with several radio interfaces for Bluetooth, WiFi, Universal Mo- bileTelecommunicationSystem(UMTS),LongTermEvolution(LTE) (amongothers)themainissueishowtotakeprofitfromthismul- tihoming opportunity in order to develop improved WBAN han- doverpracticestomaintainhealthcareservicecontinuityandcope withmobilityandubiquitouscoverageissues[5].

http://dx.doi.org/10.1016/j.comcom.2016.01.007 0140-3664/© 2016 Elsevier B.V. All rights reserved.

Eventhougheachtechnologyhasitsowncharacteristics(WiFi:

high-data-rate,shortrange,low mobility;3G:low-data-rate, high mobility,and LTE: low latency, highthroughput, high speed and Limitedcoverageoutsideurbanareas)themajorityofWiFicover- ageareasoverlapwith3G andLTEones.Furthermore,there exist widedifferencesindatageneration rateanddelay/loss-tolerances amongst the data packets generated by heterogeneous WBAN devices. For example, some low-data-rate medical sensors like heartbeat,bloodpressure,electroencephalogramsensorsmaygen- erate very time-critical data packets, which must be communi- catedto the medical staff within a guaranteed end-to-end delay deadline.Incontrast,some high-data-ratesensors(e.g.,streaming ofECG signals) may allow a certain percentage of packet losses.

Hence,an efficientdecision-makingalgorithm is neededin order tochoose thebest networkamong theavailable listaccordingto the application’s requirements, and seamlessly perform the han- doverprocess.

In the literature, severalmechanisms (reviewed in Section 2) havebeenproposed tomitigate thehandover problem. However, tothe best of ourknowledge, no effective solutionhas yet been proposedsofarparticularlyforWBANs.

The maininnovative contributionofthispaperis theproposi- tionand analysis ofa handover protocol that well adapts to the remotehealthcaredomainandtakesaccurate accountofthecon- stantevolution ofpatient’sQoSrequirements.Infact,it isimpor- tanttounderline thatmanychanges canhappen intermsofQoS requirementsofWBANsensorsovertime.Inotherterms,thesam- plingrate,therequiredlatency,andtheimportanceofsenseddata, which are key metrics for network selection, can indeed change accordingtothepatient’scondition. Forthisreason,inMADMHA, the handover triggering process depends on the patient condi- tions. Furthermore, the proposed handover algorithm allows the patienttoremainconnectedtothecurrentnetworkthatstillguar- anteestherequiredQoSoftherunninghealthcareapplicationeven thoughanewnetworkisdiscovered.Itisalsobasedonasofthan- doverapproach,whichisreferredtoas“make-before-break”,thus ensuringthatthemobilepatient’snewconnectioniscreatedatthe targetPoAbeforetheoldPoAconnectionisreleased.

Insummary,ourpapermakesthefollowingkeycontributions:

• Weconductatechnicalandnumericalcomparisonbetweendif- ferentwirelesstechnologies(i.e.,IEEE802.11x,UMTS,LTE).

• Wedefine asetofhealthcaremonitoringapplications(ortraf- ficcategories)torepresentgeneralmonitoringtrafficdata,high priorityandemergencydata.

• Wedeterminefortheconsideredapplicationstheircorrespond- ing QoSmetrics,whichareused byourhandoverapproachto performanoptimalnetworkselection.

• We propose a novel and effective multi-attribute-decision- making handover approach which guarantees the required quality of servicelevel while taking accurate account of net- workhistoryanduserpreferenceissues.

• We performathoroughperformance comparisonbetweenour proposed approach andexisting ones. Numericalresults show that MADMHA is indeedeffective; it significantly reducesthe packet overhead and handover frequency, while limiting the packetlossratio.

The paperisstructured asfollows:Section 2discusses related work. Section 3 highlights the main characteristics ofLTE, UMTS andWiFi technologies. Section 4 introduces our WBAN network modelandtrafficcategories.InSection5wepresentourMADMHA algorithm,while we illustrate and discussnumerical results that showtheefficiencyofourproposalinSection6.

Finally,Section7concludesthispaperandpresentssomefuture works.

2. Relatedwork

Since in our previous works [6–8] we already addressed the intra-body communicationlevel,we focushereonthe extra-body one.

Specifically, we consider a mobile WBAN-based environment andtacklethehandoverissue,whichisverychallenging[5,9–12]. Therefore, in this section we survey several recent works deal- ingwiththehandoverphenomenonthataretightlyrelatedtoour work.

Even though in all these works the target is the same (i.e., keeping mobile devices always connected), handover decision parameters vary from one proposal to another. The handover decisionmaydependononeoracombinationofstaticparameters (power consumption, monetary cost, security) and/or dynamic ones(data rate, available bandwidth,RSSI, Signal-to-Interference- and-Noise Ratio (SINR), latency, velocity, user preference and so on.) [13]. The most widely used parameters in single metric handover approaches [14] are the RSSI and the Data Rate (DR).

The RSS-VHD approach (RSS-based Vertical Handover Decision) proposedin[14]isbasedonacomparisonbetweenthemeasured RSS value by different mobile terminals and the defined RSS thresholds.Therefore,whentheRSSofWLANdropsbelowdefined thresholds, the registration procedures are initiated for Mobile terminal’s handover to the 3G network. On the other hand, the DR-VHD approachproposed by thesameauthors isbased onthe evaluationoftheoffered datarateinthecurrentservingnetwork with respect to available, candidate networks. A DR-VHD takes place whenever a transition to a candidate network can ensure a datarategain.Althoughsinglemetrichandover approachesare easy to implement, they can suffer from lack of effectiveness.

Therefore,unnecessary handover mayoccur, andthiscan lead to highenergyconsumption andhugepacket lossduetoping pong effect. Authors in [15] have combined several parameters, such as security, energy consumption, bandwidth and link qual- ity in a normalized weighted cost function to select the best available network. However, evaluating the security level of a network via a simple parameter in a cost function is a diffi- cult task. Network history and traffic differentiation are also overlooked in this latter work. Khan et al.[16] use a cost func- tion to perform the handover decision between WiMAX and WiFi. In a double coverage area, users switch to the network having the smallest cost. However, the cost function adopted in this work is the sum of heterogeneous and non-additive parameters.

Authors in[17]makeuseofGameTheorytoaddressthehan- doverproblem.Indeed,thebesttargetnetworkismodeledthrough aBayesianNash-equilibriumpointthattradesoff between quality of servicemaximization and cost minimization. Even though the proposalensureslowhandoverdelayandcommunicationprices,it completelyoverlookstrafficdifferentiation.

Other works have considered a cross-layer approach [18,19]. In[18],across-layerhandovermanagementframeworkisproposed and handover triggers are launched according to an information databasegathered fromdifferent layers. Then, two types ofhan- dover are defined: imperative and alternative handover. A han- doverisconsidered asimperativeifthereisasignalstrengthloss.

Ontheother hand,itisconsideredanalternativeonewhenthere isaneedforQoSenhancementduetoachangeinapplicationre- quirementsand/oruser’spreferences.

Similarly,Rehanetal.[19]intheirapproachgatherinformation fromtheMediumAccessControl(MAC),transport,andapplication layersforhandovertriggering,consideringuser’spreferences.

According to their requirements, users may choose a cost ef- fective network or the best performing one even if it is costly.

Table 1

Properties of WiFi, UMTS and LTE [28] .

IEEE 802.11a IEEE 802.11b IEEE 802.11g IEEE 802.11n UMTS LTE

Radio OFDM OFDM OFDM OFDM CDMA OFDM

technology MIMO MIMO MIMO MIMO FDD SCFDMA

Frequency 5 GHz UNII 2.4 GHz 900 MHz 2.4, 5 GHz 450, 850 MHz, 80 0, 180 0 MHz 1.9, 2,

2.5, 3.5 GHz

Range ∼30 m ∼30 m ∼30 m ∼50 m ∼26 km 30–50 km

Peak 54 11 54 600 14.4 326.4

downlink (Mb)

Peak 54 11 54 600 0.3840 86.4

Uplink (Mb)

Typical downlink 20 5 20 – 2 –

throughput (Mb)

Access protocol CSMA/CA CSMA/CA CSMA/CA CSMA/CA WCDMA OFDMA

Network Infrastructure Infrast. Infrast. Infrast. RAN RAN

topology

Main Short Short Short Short Handover Expensive

disadvantages range range range range problems

However, eventhough avariety ofparametersfromdifferentlay- ersareusedtoperformthehandover,thedecision-makingprocess described insuch a work is not well-defined and somehowam- biguous.

MADM is a process for making preference decision over the available alternatives. Simple Additive Weighting (SAW), Weight- ing Product (WP), and Technique for Order Preference by Sim- ilarity to Ideal Solution (TOPSIS) are existing MADM methods that deal with the issue of selecting the best one from a set of alternatives, according to multiple attributes. These differ- ent methods are adopted by researchers to treat the handover problem [20–22]. The authors in [20] combine SAW and WP to rank the listof visitednetworks andselect thebest one. TOPSIS hasbeenusedin[21]toensurea seamlesshandover.In[21],the authorspointouttheworstandbestnetwork,anddefinethetar- get network as theclosest to the idealsolution and the farthest fromtheworstone.

AlthoughtheseMADMmethodshavemanystrengths,theysuf- fer from some limitations. Indeed, according to [22] both SAW and WP suffer fromthe lack of accuracy in identifying the rank of alternatives, while TOPSIS suffers fromthe ranking abnormal- ityproblem.Such problemoccurswhenalow rankingalternative is removedfromthe list ofcandidates (e.g.,one network is sud- denly disconnected),so that the order ofhigher ranking alterna- tiveswillchangeabnormally.Thereby,toimproveMADMforseam- lesshandover,weproposeinthisworkan hybridapproach,called MADMHA,that leverages thestrengths and overcomesthe weak- nessesofexistingMADMmethods.

3. Comparisonofwirelesstechnologies

A wide variety of different wireless technologies exist (i.e., WiFi, WiMAX, UMTS, LTE, LTEAdvanced), some in directcompe- tition with each other, others designed for specific applications [10,23–27]. In this section, however, we focus on WiFi, LTE and UMTS for providing extra-body communication since nowadays commercialized personal devices are equipped with radio inter- faces using these wireless technologies, in addition to Bluetooth for intra-body communications. For an exhaustive review of dif- ferent wireless technologies, with their advantages and disad- vantages, forpromoting mobile e-health applicationsplease refer to[10,27]

WLANs, oftenknownby theircommercialproduct nameWiFi, have a variety of standards, each represented with a letter suf- fix. Ofthese,the standardsthat aremost widelyknown areIEEE 802.11a, IEEE 802.11b, IEEE 802.11g, and IEEE 802.11n. Besides,

4GWAN technologies(e.g.,LTE) provideeven higherbitratesand many architectural improvements than 3G ones (UMTS). These technologiescanbe evaluated bya varietyofdifferentmetrics of which some are described in this section. Therefore, Table 1 il- lustrates a rough comparison between different WiFi standards, UMTS,andLTEtechnologies.

Thecomplementaryofthesetechnologiesintermsofcoverage area,technicalcharacteristicsandcommercialopportunitiesforthe operators fostered the developmentof smartphones andPDA in- tegratingmultipleradio interfaces.Theemergenceofsuchmobile terminalsequippedwithvariousinterfacesprovidesmanyinterest- ingbenefits,suchaspermanentandubiquitousaccess.However,it raises vertical handover issues which is the process of handover thatoccursbythemovementofamobilenodeamongthehetero- geneoustechnologies[16].

4. Networkmodelandtrafficcategories

Despitethefact thatWBANshavemanypromisingapplication domains, thiswork focuseson the remote healthcare monitoring one.Therefore,inthissectionwe presentourhealthcare network modelaswell astheproposed traffic categorization inthislatter domain.

4.1. Networkmodel

Inournetworkmodel,weconsiderasetN ofWBANs(e-health users)andasetMofheterogeneouswireless networks.Wefocus inthisworkonremotehealthcaremonitoringapplications,andwe definethree categoriesoftraffic(GM,DS,andEM) whicharede- scribedindetailinthenextsection.Letusdenotethesetofappli- cations(ortrafficcategories)byK=

{

^GM,DS,EM

}

^.

Asmentionedbefore,aWBAN’scollecteddataistransmittedto its final destination via a Personal Device (PD). We assume that thePDis equippedwithan IEEE802.15.6 interfaceforintra-body communication(communicationbetween thePD andthesensors ofthesameWBAN)andotherradio interfaces(like WiFi,3G, 4G, etc.) to ensure extra-body communication (communications be- tweenthePDandthemedicalstaff).

AWBAN(orapatient)maymoveindifferentplaces,likeamo- bile system, and wants to ubiquitously disseminate its collected data to the e-health staff (doctor, nurse, emergency car, and so on).Hence, an efficiente-health systemshouldget aWBAN con- nectedanywherehegoestakingintoaccountQoSe-healthapplica- tionrequirements.Otherwise,itshouldallowthePDtobealways

Fig. 1. Network model architecture.

Table 2

Delay and bit rate requirements of healthcare data [28] .

Data source Bit rate [bps] Delay [s] Sampling rate [Hz]

Electrocardiogram (ECG) 10 –100 k < 10 63 –500 Blood pressure [mmHg] 10 –30 > 120 63 Non-invasive cuff 0.05 30 –120 0.025 Cardiac output [ L / min ] 1k < 10 63 CO 2concentration 1k 30 –120 63 Temperature ( ^◦C ) 0.3 > 120 0.02

connectedto the bestavailable network mainly in multicovered zones.

Fig.1illustratesournetworkmodelarchitecture.

4.2.Trafficcategories

Unlike conventional Wireless Sensor Networks (WSNs), each sensornode inaWBANhasitsownrequirementsintermsofde- lay,prioritylevelanddatarate.Forinstance,inmostreallifetrials, ECGsensorreadingsareconsideredwithhigherprioritythanSpO2 (percutaneousoxygensaturation)andbodytemperaturedata.ECG needsa highsamplingratewhilea fewsamplesper dayaresuf- ficientforSpO2andbodytemperaturesensors.However,itisim- portant to underline that many changes can happen in terms of QoSrequirementsofWBANsensors overtime.Thesamplingrate, the required latency, and the importance of sensed data change accordingtothe patient’scondition. Forinstance,inpost-surgery conditions,patientsrequirecontinuous24-hmonitoring.However, forinjured andstablepatients, thetest frequencymay be setto everyhourorevenless.Furthermore,whenreadingscomingfrom alow priority sensor(SpO2, pulserate, temperature,etc.) exceed normalthresholds(e.g.,bodytemperature≥40^◦C)thesensor’spri- orityisincreased.Table2showsindetailtheheterogeneouschar- acteristicsofsomecommonlyusedmedicalsensors.

Thereby,to guaranteehealthcareserviceefficiency,itis neces- sarytodesignWBANprotocolsthatcanhandlesuchheterogeneity andthevariationofQoSrequirementsovertime.

Accordingly,theIEEE802.15.6standard[29]hasspecifiedauser priorityfieldintheWBAN’sframestructure.

Furthermore,apacketsubtypefieldallowsuserstodefineother datasubtypes.Therefore,inview ofthat,inoursolution wesug- gest the following mapping between traffic designation and the packetsubtypefield:

• GeneralMonitoring(GM)packetsforordinarymedicaldata;

• DelaySensitive(DS)packetsforhigh-prioritymedicaldata;

• TheMandatoryEmergency(EM)datasubtype.

Moredetailsondatapackettypeclassificationandfieldencod- ing arepresentedinTable3.Accordingly, togive meaningtothis classificationmade,inthenextsection,weproposeahandoverap- proachawareoftheheterogeneousandtimevaryingQoSrequire- mentsofeachapplicationtype.

5. SeamlessMulti-AttributeDecisionMakingHandover (MADMH)approach

Havingpresentedthe network modelandtraffic categories,in thissection we introduceour MADMH approach.The mostchal- lengingissuesthatourapproachwillencounterare:

• howto ensurea seamless switchingfromone networkto an- other(better)one;

• how to satisfy QoS requirements fordifferent WBAN applica- tions;

• howtoreducethenumberofhandover.

Our handover process can be divided into two main phases:

(i) network monitoring andinformationgathering phase, and(ii) decision-makingandhandoverexecutionphase.

Beforedescribingindetailthesetwophases,letusfirstpresent theQoSattributes,correspondingtothepreviouslyintroducedap- plications,usedbyourapproachtochoosethemostefficientnet- work(ortechnology)amongsttheavailableones.

Table 3

Data packets type classification and field encoding.

User Traffic Packet Packet Description

priority designation type subtype

7 Emergency or event report Data EM They are the most critical data packets. They should be forwarded in brief time and reliable way. They are used to report data from important sensors, like ECG, EEG, and all sensor readings that exceed normal thresholds 6 Medical data Data DS The video traffic type defined by the standard is designed for non-medical

applications, e.g., video gaming. Moreover, some medical sensors like EMG and motion sensors generate delay sensitive data. For this reason, we defined the DS type as the medical data that must be delivered within a stringent deadline and tolerates a reasonable packet loss

3 Controlled load Data GM The lowest priority is given to GM packets. They correspond to regular measurements of patient physiological parameters that typically indicate normal values received from lower priority sensors (SpO2, Non-invasive cuff, etc.)

5.1. QualityofServiceattributes

QoS attributes are chosen according to traffic categories pre- sentedinSection4.2andconsistoftheDataRate(DR),thePacket DeliveryRatio(PDR), theReceivedSignalStrengthIndicator(RSSI) andthePointtoPointpacketDelay(PPD)whichrepresentstheto- taltransmissiontimethatisneededtosendadatapacketfromthe PDtothePointofAttachment(PoA).Inthisworkandtheperfor- manceevaluationsection, theseQoSmetricsare computedbased on theexpressions detailedhereafter,in linewithexisting litera- ture.

Inparticular,theRSSIindBmcanbeexpressedasfollows:

RSSI_dBm=Pt−10

η

^log_d^d

0−X, (1)

wherePtrepresentsthetransmissionpowerofthesender,

η

^isthe

pathlosscoefficient,disthedistancebetweenthesenderandthe receiver, d₀ is a reference distanceto be in a far field condition, andXisarandomvariablethattakesintoaccountfadingeffects.

The basic notations used in this paper are summarized in Table4.

ThePDRisgivenby:

PDR= Numberofreceivedpackets

Numberofsentpackets ⁽²⁾

We assume that, in a WiFinetwork scenario,the PD accesses themediumusingtheCSMA/CAMACprotocol,theaccessinUMTS is based onWideband Code Division Multiple Access (W-CDMA), while itis based onSC-FDMA/OFDMAin LTE.Therefore, thePPD inWiFi,UMTSandLTEscenariosaregiven,respectively,byexpres- sions(3),(4)and(5)[30]:

PPDW iFi=T_{di f s}+T_{si f s}+T_{bo f f} +T_data+T_ack, (3) wherethedatapackettransmissiontime T_dataisequaltothesum oftransmissiondelayandpropagationdelay:

T_data= T R+d

S

andtheglobalsizeT ofatransmittedpacketis:

T =PHY_hl+MAC_hl+MAC_{f t}+P_load

PPD_UMTS=5ms+X·10ms+

α

^l

^{·10ms, (4)}

PPDLTE=T_up+Bu f_t+R_d+U_sch_R+U_sh_g

+UE_d+eNodeB_d, (5) 5.2. Networkmonitoringandinformationgathering

Asafirststep,informationgatheringmaysignificantlyinfluence the success ofthe overall handover process. Infact, the network

Table 4

Notations and definitions.

P t Transmission power of the sender d Distance between the sender and

the receiver

d 0 Reference distance to be in far field condition X A random variable that takes

into account fading effects η Path loss coefficient

(for the free space case η= 2 ) T difs Distributed interframe space time T sifs Short interframe space time

l Payload length of the transmitted packet

α Length-factor

T ack Time to receive an acknowledgement T boff Backoff time

T data Data packet transmission time R Network data rate

S Speed of light PHY hl Physical header’s length MAC hl MAC header’s length MAC ft MAC’s footer P load Packet payload

T Global size of the transmitted packet T _ up LTE uplink transmission time U _ sh _ R LTE uplink scheduling request delay U _ sh _ G LTE uplink scheduling grant delay R _ d LTE retransmission delay eNod eB _ d LTE eNodeB processing delay Bu f _ t LTE buffering time

UE _ d LTE user equipment processing delay

mustbewellmonitoredandthehandoverdecisionmustbemade atthe right time to establish a right new connection before the servicedisruption ofthe current one.Hence, information gather- ing process must be invokedin compliance with the application requirements,network conditionsand user preferencevariations.

Therefore,ourinformationgatheringtriggerscanbeoftwotypes:

user-preferencetriggers andQoStriggers.Theinformationgather- inguser-preferencetriggerreferstoachangesetbytheuserinhis networkpreferencelist.Ontheotherhand,informationgathering QoStriggers areintheirroletwo-folds:applicationlayerandlink layertriggers.Thefirsttypereferstoachangeintheapplication’s QoSrequirementsduetoachangeinitstype;forexample,dueto anaggravationofthepatient’ssituation,theapplicationtypemay changefromageneralmonitoringapplicationtoanemergencyone andviceversa.

Theinformationgatheringlinklayertriggersaredrivenbylinks going down orup. In particular, they depend on monitored RSSI valuesvariations. Wemake useof RSSI, duetothe factthat it is hardtoestimateothermetricslikeSignaltoInterference(SIR)and

SignaltoInterferenceandNoise(SINR).Furthermore,theinterfer- enceissuewillbeconsideredinafuturework.

Infact,alinkgoingdowntriggerisfiredwheneverthePDde- tects that the RSSIis not appropriate forthe current application type.Forthelinkuptrigger,itisfiredwheneverthePDdetectsa newnetwork witha high RSSIvalue. Therefore, adaptivethresh- oldstocontrol the RSSI(andsimilarly theother metrics)are de- fined andset according to applications types.Indeed, inour ap- proach,for each application type ∈{GM, DS, EM} andeach QoS evaluation metric ∈ {RSSI, PDR, DR, PPD} we define a min and amax threshold. Forexample,the min andmax RSSIthresholds that we use for GM applications are respectively −87 dBm and

−65dBm. All usedadaptive QoSmetricthresholds are presented belowinTable6.

Tosummarize,fourscenariosmaypromptthePDtotriggerthe informationgatheringprocessandafterthedecisionmaking:

• Scenario 1: The user selects another network since hefavors thislatterwithrespecttothecurrentone.

• Scenario 2: An important change in the captured data values occurs, whichleadsto achangeof theapplicationtype either fromGMtoEMorfromEMtoGM.

• Scenario 3: A linkdeterioration is detected via the controlled RSSI values (due to WBAN’s distance to a PoA, an obstacle and/orcongestion).

• Scenario4:AnewnetworkisdiscoveredduringWBAN’smove- ment(capturingasignalwithahighRSSIvalue).

Hence, whenever a trigger among those just mentioned is launched, the information gathering process of the current net- workstarts.Ifthe currentnetwork stillmeet thecurrentrunning applicationrequirements,thePDkeepsconnectingtoit.Otherwise, it begins collecting information about the list of available net- works.The gathered information relatedto the user preferences, currentrunning application and its QoS requirements aswell as thelistofavailable networksandtheir correspondingQoSparam- eters(RSSI,PPD,DR, PDR)willbe communicated toMADMHAto maketherighthandoverdecision.Thedecisionprocessiswellde- tailedinthefollowingsection.

5.3.Multi-AttributeDecisionMakingHandoverAlgorithm(MADMHA)

ThefundamentalobjectiveofMADMHAistodetermine,among afinitesetMofavailable wirelessnetworks, theoptimalonefor anapplicationk∈K; theselectednetwork mustprovidethebest availableQoSandnetworkstabilityaccordingtoapplicationk’spri- ority. MADMHAdecisionattributesmay be groupedintwo main categories:QoSrelatedanduserrelated.

The MADMHA algorithm consists of the following five main steps(detailedinSections5.3.1–5.3.5):

5.3.1. ConstructionoftheQoSmatrix

The main target here is to choose among the several avail- ablenetworksM,havingmdifferentQoSparameters,theonethat meetsthebestanapplicationk∈K’sQoSrequirements.Hence,for each applicationk, we model thisselection problemwith a QoS decisionmatrixcalledU_QoS^k ,definedasfollows:

U_QoS^k =

⎛

⎜ ⎝

U₁₁^k U₁₂^k ... U₁^k_m U₂₁^k ... ... U₂^k_m

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

U_|^k_M|1 ... ... U_|^k_M|m

⎞

⎟ ⎠

where m is the number of QoS attributes (in our case m˜=4) andU_{i j}^k representsthe utility ofthePD when choosing thealter- nativewireless networki withrespect to the QoSattribute j for

i∈

{

¹,...,M

}

^{and j}∈

{

¹,...,m

}

^{.TheutilityfunctionU}i j^kcanbeex- pressedasbelow:

U_{i j}^k=

⎧ ⎪

⎪ ⎨

⎪ ⎪

⎩

−1, V_{i j}^k<V_min^k C ^V

k i j−V_min^k

Vmax^k −Vmin^k , V_min^k ≤V_{i j}^k<V_max^k 1, V_{i j}^k≥V_max^k

(6)

whereV_{i j}^kisthecurrentvalueoftheQoSattributejinthealterna- tive networki.V_max^k denotes theidealupperthresholdoftheQoS attribute j for an application k. V_min^k is the minimal requirement valueofj QoSattributethat thek−thapplicationrequires.Cisa coefficient that reflectsthe relativechange ofthej attribute over time,andisusedtopromotealternativeimprovingnetworks.This latteriscomputedasfollows:

C=

⎧ ⎪

⎨

⎪ ⎩

1, t=0

1+V_{i j,t}^k −V_{i j,t−1}^k

V_{i j}^k_,t , t≥0 ⁽⁷⁾

withtrepresentingtime,V_{i j}^k_,t the current value ofthe QoSattribute jinthealternativenetworkiandV_{i j,t−1}^k thelastmeasuredvalue(at timet−1)ofthislatter.

5.3.2. ConstructionoftheglobalQoSvectorV

After modeling theQoS parameters ina QoSMatrixU_QoS^k , the second step consists in multiplying the columns of each row of thislatter matrixto obtain aglobal QoSvector VG^k. VG^k may be expressedasfollows:

VG^k=

⎛

⎜ ⎝

VG^k₁ ...

...

VG^k_|M|

⎞

⎟ ⎠

where VG^k_i =

δ

∗

m

j=1

|

^Ui j^k

|

⁽⁸⁾

δ

⁼

(

⁻¹

)

^{NNbb+2+1 (9)}

Note that, it is quite possible that one network i has more than one QoS parameter j not meeting the QoS requirements of anapplicationk(^U_{i j}^k=−1)^{(seeEq.(6)).WenotebyN}bthenum- berofthatQoSparametersvaluesthatarelowerthantherequired thresholdsinthealternativenetworki(^U_{i j}^k=−1)^.IfNbisaneven- numberedinteger,theresultingVG^k_i valuewillbepositive,evenif network idoesnot meetall the QoSrequirementsofthecurrent running application. To copewith thisproblem, we make useof the

δ

^{coefficient,whichguaranteesthat:}

∀

^networkii∈

{

¹,M

}

;VG_i<0⇔N_b≥1

Proof. ForN_b=0:

^NN^b_b⁺+²1

=

²1

=2.So,

δ

=(−1)²=1. ForN_b ≥1:

^NN^b_b⁺+1²

=

(¹+_N¹

b+1)

=1.So,

δ

=(−1)¹=−1. Therefore,

∀

^networkii∈

{

¹,M

}

;VG_i<0⇔N_b≥1.

5.3.3. ConstructionofthefilteredQoSvectorV_f

In order to keep inthe list of candidatenetworks only those havingsufficientQoSservices,wefiltertheVGvectorasexplained below. First, we compute the global minimal QoS requirement valueVG^k_min,where

VG^k_min= m

j=1

U_{i j}^k_,min

U_{i j,max}^{k (10)}

Fig. 2. The handover process flowchart.

Second,wediscardnetworkshavingaglobalQoSvaluesmaller thantheminimalrequirementVG^k_min fromtheVGvectortoobtain thefilteredQoSvectorV_f.Hence:

VG^k_i ∈V_f⇔VG^k_i >VG^k_min

5.3.4. ConstructionofthenormalizedweightedvectorV_N

Inourproposedmechanism,weconsidertheuser’spreference, whichmayberelated,forexample,tothenotionoffamiliarityand confidenceina givennetwork, orits monetarycost.Therefore,in orderto geta realistic evaluationofdifferentnetworks, we com- bine in our evaluation QoS-related parameters with user-related ones.AstheQoSparametersaretakenintoaccountintheV_fvec- tor,wemodeltheuserpreferenceswithrespecttocandidatenet- workswithavectornamedUP.Finally,wesumthetwovectorsto obtainanormalizedweightedone,calledV_N.

VN=

α

⎛

⎜ ⎝

VG^k₁ ......

VG^k_|M|

⎞

⎟ ⎠

⁺

β

⎛

⎜ ⎝

UP1

......

UP_|M|

⎞

⎟ ⎠

⁽¹¹⁾

Here,

α

^and

β

^{are theweighting coefficientsassigned respec-}

tively to the QoSand user preference vectors, while

|

M

|

^{is the}

numberofcandidatenetworks.

5.3.5. Selectionofthebestsolution

Thefinaltaskistoselecttheoptimalsolutionwhichisthenet- workihavingthegreatestV_N,i.

5.4.Handoverexecution

Oncethedecisionismadeandanewtargetnetworkischosen, thehandoverexecutionwillbe carriedout.Accordingto[31],the handoverexecutiontechniquescanbeoftwotypes:

• Hard handover:The currentnetwork isreleasedandonlyafter thisoperationthetarget oneisengaged.Withsuchtechnique, thehandoverisdoneinabreak-before-makeway.

• Softhandover:Inthiscase,theconnectiontothetargetnetwork is established before the connection to the source is broken, hencethislattertechniqueiscalledmake-before-break.

As mentioned before, in a e-health monitoring application, treated data is indeed crucial, and the slightest mistake can se- riouslyaffectpatientssafety.Thisrequiresservicecontinuitydur- ingthehandoverexecutionprocess,whichisourmaingoal.Inthe caseofahardhandover,aservicedisruptionperiodcanbeexperi- encedduetothetime neededfordisconnectionfromthecurrent cellandconnectiontothetargetone.However,thisriskisavoided inasofthandovercase.Inourapproach,wehenceadoptthesoft handovertechnique.

Tosummarize, the proposed seamless Handoverapproachcan be outlined in the flowchart depicted in Fig. 2. In fact, asmen- tioned before, the handover process may be fired either with a useroraQoStrigger.Inthefirstcase,thetarget networkiseval- uatedby computingitsVGvalue. IftheobtainedVGis greateror equaltotheaveragerequiredQoSvalue,VG_min,theuserconnects

Table 5

Simulation parameters.

Wireless Max Tx Cell Data Recv_signal

technology power (dBm) radius (m) rate (M) threshold (dBm)

UMTS 43 600 2 −121

IEEE 802.11b/g 20 100 4 −91

IEEE 802.11b 27 100 11 −89

IEEE 802.11a 17 75 6 −89

IEEE 802.11g 26 75 6 −92

IEEE 802.11n 18 125 24 −72

LTE 46 10 0 0 13 −123 . 4

tothislatternetwork,otherwiseheremainsconnectedtothecur- rentone, hencelimitingtheping-pongeffect.Inthesecond case, theuser looks fora QoS enhancement.Thus, the system gathers informationofallavailablenetworks,executestheMADMHAalgo- rithm,andifabetternetworkexists,thenconnectstotheselected one.

6. Performanceevaluation

Thissectionpresentsnumericalresultsthatillustratethevalid- ityof theproposed MADMHAalgorithmfornetwork selectionby mobileWBANs.Morespecifically,wetestthesensitivityofthepro- posedalgorithmtodifferentparameterslikethenumberofWBANs andtheirtraffictypes(EM,DS,andGM).

Hereafter, we first describe thesimulation setup andthen we analyze and discuss the performance achieved by the proposed approach.

6.1.Simulationsetup

The proposedapproachandalgorithmsareimplemented using theOMNeT++ simulator[32].The deployednetwork iscomposed of10mobileWBANsinasimulationareaof1000×1000m².This areaiscoveredbyeightWiFiaccesspoints(APs),four3Gbasesta- tionsandfourLTEeNodeBs.Accessnetworkselectionrelatedinfor- mationforeachoftheconsiderednetworksisreportedinTable5. Mobile WBANs move in a pedestrian environment (i.e., at a 0.5 m/s speed) according to the random waypoint model [33]. Moreover, the AP, eNodeB and BS positions are unknown a pri- ori. Consequently, the probability to perform a handover from one network to another, is strictly dependent on the network characteristics. In our simulations we consider three different healthcareapplications:EM,DS,andGM.Inparticular,EMpackets correspondtoECGandEEGreadings,whileDStrafficisgenerated byMotion sensingandEMG sensors. GlucoseandBody tempera- turesensors’ readings are considered as generalmonitoring traf- fics. However, assaid before, we note here that traffic priorities mayvarydependinguponthevaluesgeneratedbythesensors.For instance,bodytemperaturereadingsmayproduceEMtrafficflows iftheirvalues exceedcertain thresholds.QoSparameters, aswell as

α

^and

β

^{coefficients used in our} simulations, are reported in Table6,andtheyaredefinedincompliancewiththerealsensors requirementspreviouslyreportedinTable2.

For user preference parameters, we consider two parameters whicharethemonetary costandthepatient’sconfidenceincan- didatenetworks.

6.2.Simulationresults

Simulationsareperformedinordertovalidatetheeffectiveness of the MADMHA algorithm, in comparison with that of existing (reviewedinSection2),RSS-VHDandDR-VHDapproaches[14].

Table 6

QoS parameters of different application types.

EM DS GM

Min delay threshold (s) 8 10 80 Max delay threshold (s) 10 30 100

Min PDR threshold (%) 97 80 70

Max PDR threshold (%) 100 100 100 Min RSSI threshold (dB) (−50) (−65) (−87) Max RSSI threshold (dB) (−40) (−55) (−65) Min data rate thresholds (kbps) 290 330 30 Max data rate thresholds (kbps) 400 600 100

α^{coefficient 3 2 1}

βcoefficient 1 2 3

Fig. 3. Handover frequency per hour for different handover approaches.

6.2.1. Numberofhandoverevaluation

Sincethelimitationofping-pongeffectisamandatorytask,we considerthenumberofhandoverasanimportantmetrictoprove theeffectivenessofourMADMHAalgorithm.

Therefore,toevaluatethismetric,weassumethat eachmobile WBANisbydefaultassociatedtoaWiFiAP,andthisAPbelongsto the cheapest and mostsecure network (user preferred network).

Additionally,atthebeginning,weassumethatallWBANsarerun- ning a GM application (controlling the body temperature of the patient).However,throughoutonesimulatedhour, thebodytem- peratureofsomeWBAN(patient) risestoreach40^◦C,hence,the runningapplicationisnolongerconsidered asGMbutasEM.Af- tera quarterofanhour,thetemperaturedropstoanormallevel, whichleadstoanovelchangeoftheapplicationtypefromEMto GM. While patientsare doing their routine activities, they move andmaybefaraway fromtheirdefaultAP(i.e.,gosomewherein the corridors,media room, orinthe playground, forexample)as wellastheymaydiscovernewWiFiAPswithhighRSSIvalue,de- pendingontheirlocation.Inthesecases,tokeepthepatientbest connected,hewillhavetoswitchtoanothernetwork(WiFi,UMTS orLTE)thatsatisfiestheQoSoftherunningapplicationtakinginto consideration the user preference. In other terms, the four han- dovertriggersdescribedinSection5.2areusedinthisscenario.

Therefore, Fig. 3 shows the average number of handover per hour experienced with MADMHA, RSS-VHD, and DR-VHD during aWBAN’smovement.NumericalresultsprovethatMADMHAper- formswell compared toRSS-VHDandDR-VHD, andstronglylim- itsping pongeffects.Thisisduetothefactthat, unlikeRSS-VHD and DR-VHD, MADMHA is based on a combination of different QoS, user preference and network history parameters. Moreover, thehandoverisexecutedonlyifneeded.Indeed,ifthecurrentnet- workstillmeetsthecurrentrunningapplicationrequirements,the handoverdoesn’ttakeaplace,evenifabetternetworkexists.

Fig. 4. UMTS use percentage in a WiFi covered open space versus the number of users and traffic types.

6.2.2. Percentageuseofwirelesstechnologies

SelectingthebestnetworkisamandatorytaskforWBANs,es- peciallyforcriticaltraffic.Therefore,inthenextsimulationscenar- ios, we assumethat WBANusersare locatedina dedicatedopen spacecoveredbyWiFi, andwe increasethenumberofconnected usersperAP.Thisopenspacemaybeawaitingroominahospital, aseminarorconferenceroom,wherepatientscanattendahealth- awarenessconference.Therein,thenumberofpatientsisnotfixed:

theycanmovearound,enteronebyoneoringroups,thusincreas- ingthechannelcontention.

Scenario 1. In this scenario, we assume that the available net- worksareIEEE802.11b/gandUMTS.

We measure thefrequencyuse ofUMTSforthe three consid- ered traffic types versus the number of WBAN users located in the open space covered by WiFi. Obtained results are illustrated inFig.4.

WeobserveinFig.4that,foralltraffictypes,thepercentageof UMTSconnections linearlyincreaseswhile increasingthenumber ofWBANusersintheroom.Wealsonotethatsuchanincreaseis stronglyrelatedtothetraffictype.Morespecifically:

• for10users,allpatientsuseWiFi;

• for20users,allpatientsgeneratingDSandGMtraffickeepus- ingWiFi,while16%ofEMusersmigratetoUMTS;

• when the numberof indoor users reaches40, the percentage ofmigrationtoUMTSforGM,DSandEMpeak,respectively,to 29%,38%and82%.

These results can be explained by the fact that, increasing the numberof WBANusers per APleadsto network quality and resources degradation. On the other hand, EM traffic requires a higher network performance than DS and GM traffics. So, even thoughtheUMTSnetworkisexpensivecomparedtoWiFi,itisse- lected whenthenumberofconnectionsper APexceeds 10.How- ever,theUMTSnetworkisselectedonlywhenthenumberofcon- nectedWBANsperAPexceeds20forDSandGMtraffics.

Scenario2. In thisscenario,the available networksare WiFiand LTE.InordertocomparethedifferentWiFistandards(IEEE802.11a, IEEE 802.11b, IEEE 802.11g andIEEE 802.11n), we change the AP typeineachsimulationrun.Therefore,wemeasurethefrequency useofLTEforthethreeconsideredtraffictypesversusthenumber ofWBAN usersandWiFiAPtype. Obtainedresultsare illustrated inTable7.

Table 7

LTE use percentage in a WiFi covered open space versus the number of users, used WiFi technology and traffic types.

WiFi AP

Nb users IEEE 802.11a IEEE 802.11b IEEE 802.11n IEEE 802.11g

10 0 0 0 0

20 17% (EM) 10% (EM) 0 19% (EM)

4% (GM) 0% (GM) 0% (GM) 6% (GM) 40 29 % (DS) 19% (DS) 2% (DS) 31% (DS) 71 % ( EM) 39% (EM) 5% (EM) 74% (EM) 16% (GM) 3% (GM) 0% (GM) 18% (GM) 60 29% (DS) 27% (DS) 9% (DS) 49% (DS)

71% (EM) 70% (EM) 11% (EM) 83% (EM)

Wecan observethat theresults obtainedwithLTEare inline withthoseofscenario1,forallusedWiFitechnologies.Thehigher thenumber ofconnectedusers per AP, thelower isthe percent- ageofWiFiconnections.Wefurthernotethatsuchbehaviorhighly dependson traffic andAPtypes.For thethree WiFitechnologies (IEEE 802.11a, IEEE 802.11b and IEEE 802.11g), the migration of usersgeneratingEMtrafficstoLTEnetworkbeginswhenthenum- ber of connected users per AP reaches 20. However, the degree ofmigrationdiffers fromone technologytoanother. Indeed,IEEE 802.11n performs better than IEEE 802.11a and IEEE 802.11g. For example,for60usersper APandaDSapplication, thedegree of migrationtoLTEisasfollows:

• 49%forIEEE802.11g;

• 45%forIEEE802.11a;

• 27%forIEEE802.11b;

• 9%forIEEE802.11n.

Furthermore,foraGMapplicationandIEEE802.11nanymigra- tiontoLTEisnoted.Theseresultscanbeexplainedbythefactthat (i)IEEE802.11nandIEEE802.11boffergreateravailablebandwidth andcoverageareathanIEEE802.11gandIEEE802.11a,and(ii)GM applicationsrequirelessresourcesthanDSandEM.

6.2.3. PacketOverheadevaluation

ThePacketOverhead(PO)isdefinedasthetotalnumberofcon- trolpacketsaswellasdatapackets,includingretransmittedpack- etsduetocollisions. POisasourceofenergywastingandpacket loss.Hence,thehandoverprocesshastobeperformedinaseam- lesswaywhilekeepingthePOlow.

In this scenario, we suppose that 10 patients are going for a healthy walk/runin thehospital’s green garden to promotetheir physicalactivity.Weassumethat4patientsarerunningEMappli- cations(ECGformonitoringheartbeat),3arerunningDSapplica- tions(EMG),andtheother3arerunningGM(monitoringthebody temperature)applications.TheconsideredpacketratesforGM,DS andEMaresetrespectivelyto30Kbps,340Kbpsand400Kbps.In fact,we compute theamountofthe extrapaging messages,WiFi managementframesandretransmitteddatapacketscausedbythe handoveroperationduringasimulationtimeof500sforthewalk scenario witha speed of 5 km/h. More specifically, we compute theadditionaloverheadcausedbythenetworkgatheringinforma- tion, interface switching,handover execution/failure and interfer- ences.Wenotethat,duetopatientmobility,themeasuredaverage numberofeffectivehandoverexecutedrespectivelyby MADMHA, DR-VHDandRSS-VHDduring500sisequalto1,2and5.

Fig.5showstheobtainedPOvaluesforthethreehandoverap- proaches.

Asprovenbefore,RSS-VHDandDR-VHDsufferfrompingpong effect,whichinduces a highsignaling cost(total number ofcon- trolmessages toperformthehandoverprocedure)anddisruption time.Inotherterms,thehighhandoverfrequencyinRSS-VHDand DR-VHD increases the interference of radio waves, thus increas- ingthepossibilityoflosingdatapackets.Moreover,itcausesdata

Fig. 5. Packet overhead for different handover approaches.

Fig. 6. Packet loss ratio versus traffic type for different handover approaches.

packetretransmissionsandleadstoexchangemorecontrolpackets to restore the connection between sendersand receivers. Conse- quently,theselatterapproachessufferfromasignificantoverhead.

Infact,Fig.5showsthat RSS_VHhasthehighestoverhead,while MADMHAhasthesmallestPOvalue.

6.2.4. Packetlossratioevaluation

The packetlossratio(PLR)isusedtoevaluate thenetwork re- liability, and is definedas the total number of lost data packets dividedbythetotalnumberoftransmitteddatapackets.Usingthe same parameters as in the previous PO simulation, we compute herethe PLR of the three considered applications for MADMHA, RSS-VHDandDR-VHD.Fig.6illustratestheobtainedresults.

Fig.6showsthatMADMHAhasthesmallestPLRvalue,evenfor EMtraffics.We furthernote thatthePLR isthesameforalltraf- fic typesfor the two other approaches. Thisbehavior isjustified bythefactthatsuchapproachesdonotperformtrafficdifferentia- tion.Furthermore,MADMHAensuresthebestPLRsinceitprovides thesmallestPO andhandoverfrequencyvalues,whicharetightly relatedtothePLRmetric.

6.2.5. Probabilityofhandoverfailure

Ahandoverfailuremayoccurinthefollowingcases:

• IfthetravellingtimeinsidetheservingPoAcoverageisshorter thanthehandoverlatency.

Fig. 7. Probability of handover failure versus patient speed.

• Whenthehandoverisinitiatedbutthetargetnetworkdoesnot havesufficientresourcestocompleteit.

• Whenthe handoveris initiatedanddisruptedbefore thepro- cessisfinalizedhandingovertoanothertargetnetwork.

Fromthat,itcanbeseenthatthehandoverfailuredependson themobilepatient’slocation andspeed.Infact,thetimerequired bythemobilepatienttoremainconnectedtotheservingPoAwith aspeedVmaybecomputedasfollows:

T_c=D_i

V ∗

(

¹−p f

)

⁽¹²⁾

where

D_iisthedistancetotravelinsidetheservingPoAcoverage.

pfistheprobabilityto initiateahandoverwhileremainingin- sidethecoverageoftheservingPoA.

Consequently,if thespeed V islow,the remaining connection time T_c is longer as it foresees that the mobile patient takes longertimetomovetotheedgeofthecoverageanddiscovernew PoA. However, ifthe speed is high, the delay is shorterand the probability of handover failure is higher. Hence, in this scenario thenumberofhandoverfailuresisevaluated asa functionofthe WBAN(patient)’s speed. Therefore,weconsider theprevious sce- narioandonlyvarythespeed from1km/hto8km/hduringone hourofsimulation.Fig.7showsthatinallcasestheprobabilityof handoverfailure increaseswhenincreasingthepatientspeed.Fur- thermore,itcanbeobservedthatRSS-VHDhasthehighestproba- bilityoffailure.Thiscanbeexplainedbythefactthat underRSS- VHD all WBANs (or their corresponding PDs) perform handover decisionsbased onRSSImeasurements, meaning that,ifa WBAN withahighspeedentersthecoverageareaofanewAPwhoseRSSI isabovetheoneoftheAPtowhichitiscurrentlyconnected,such WBANwillinitiatethehandoverprocedure.However,asexplained in Eq.(12),the time spent within the AP’scoverage T_c will be limitedwithhighspeed V.Consequently,the handoverprocedure will not be successfully completed promoting a call drop or de- creasingtheoverallQoS.Finally,byusingtheproposedMADMHA algorithm, which is based on multiple QoS parameters, network historyanduserpreference,thepatient’sPDremainsconnectedto thecurrentnetwork that still guaranteesthe requiredQoSofthe runninghealthcareapplicationeventhoughanewnetworkisdis- covered.Moreover,we recallthatMADMHAisbasedonsofthan- doverapproachwhichisreferredto“make-before-break”.Inother terms,themobilepatient’snewconnectioniscreatedatthetarget PoAbeforetheoldPoAconnectionisreleased.

6.2.6. AdditionalEnergyconsumption

In thisscenario, we comparethe additional energyconsumed bythethreehandoverapproaches.Wesettheaverageenergycon- sumptionpertransmittedbitforWiFi,UMTSandLTE,equalto1

μ

^J,

Fig. 8. Percentage of additional energy consumed by the handover approach.

Fig. 9. Average throughput for different handover approaches.

15

μ

^Jand20

μ

^J,respectively,while thecostduetoswitching net- work interface isequalto 1

μ

^{J[29].In fact,theadditional energy}

consumed by the handover processes during a period of time t maybecomputedasfollows:

AEng_t=EE_o

v

+EE_switch (13) where

EE_o

v

presents the additional energy consumed by the total packet overhead generated by the handover operations during a periodoftimet.

EE_switchis theadditionalenergyconsumedbytheswitching interfaceoperationsduringtheperiodt.

Fig.8 showstheaverage additionalpercentageofenergy con- sumed by a PD to conduct the handover operation during one hour ofsimulation. We canobserve that when theRSS-VHD and DR-VHD handover schemes are adopted, the PD consumes much moreenergythanwithMADMHA.Thisisinlinewithourprevious simulationresults,provingthatMADMHAreducesthepacketloss, packetoverheadandpingpongeffect.

6.2.7. Throughput

An efficient Handover approach is that which ensures high throughputdespiteterminalsmobility.Thereby,usingthesamepa- rameters asinthe previous PO simulation,we compute herethe throughputforthethreehandoverapproacheswithrespecttothe simulationtime.Fig.9showsthattheMADMHAobtainedthrough- put(bitspersecond)ishigherthanthatofRSS-VHDandDR-VHR.

This may be explained by the fact that MADMHA is able to re- mainconnectedtothesamenetworkaslongaspossible,sinceits networkselectionistunedaccordingtovariousQoSanduserpref- erence parameters. Therefore, asproven in previous simulations, MADMHAlimitsthe numberofhandovercomparedwiththetwo

Fig. 10. Average End to End Delay for different handover approaches.

other approaches.Moreover,MADMHAis basedon softhandover concept which means that buffered data are sent over both old andnewconnectionswhenahandoveroccurs.Inthisway,packet lossandretransmissionareavoided.Furthermore,wecanseethat forthethreehandoverapproachesthethroughput dropsinsome timepoints. Thisisduetotheoccurrenceofhandoverprocess at thistime.However,comparedtoRSS-VHD,DR-VHDandMADMHA appeartohaveamoderatelydownwardtrendline.

6.2.8. AverageEndtoEndDelay

ForsensitiveapplicationlikehealthcaretheshortesttheEndto EndDelay(EED),thebettertheapplicationperformance.Therefore, inthisscenariowe measure theaverage(AVG) EEDforthe three handoverapproaches.ObtainedresultsarepresentedinFig.10.

Itcan beseen inFig.10that MADMHAperforms thesmallest EED.This because, the experienced packet delaydepends on the number of handover and networks reliability and as proven be- foreourprotocoloutperformsRSS-VHD andDR-VHDin handover frequencyandPO.Infact, frequentlyoscillatingbetweendifferent networksdeteriorate the network performance leadingcollisions, interferenceandpacketretransmissions.

7. Conclusionandfutureworks

In this paper we addressed the handover issue in Wireless BodyArea Networks, focusing our attentionon three radio tech- nologies (namely WiFi, 3G and 4G access networks). In par- ticular, we proposed a hybrid Decision Making Handover Algo- rithm, named MADMHA, which ensures ubiquitous connectivity for mobile WBAN users throughout their roaming. Furthermore, MADMHA classifies the generated healthcare traffic into three classeshavingheterogeneousQoSrequirements(GeneralMonitor- ing, Delay Sensitive and Mandatory Emergency). Based on such classification,MADMHAperformsan intelligent networkselection accordingtoQoSrequirementsofeachapplicationtype,whilefur- thertakingintoaccountthenetworkhistory.Weperformedathor- oughnumericalanalysisofthe proposed algorithm,showingthat our approach achieves very good results in terms of number of handoverexperiencedbyusers,energyconsumption,packetover- headandreliability.

Threeinterestingareasforfutureworkcanbeidentifiedinthe contextofthepresentedwork.MADMHAsolutionthatwedevelop allows us to consider additional attributes and to define various relativeimportancebetweensuchattributes.Indeed,othernetwork attributes(e.g., security of the candidate network)can be added toconsider moreselection objectives.Anotherissueconcerns the setting ofweighting parameters. How can we define exactly the relativeimportance ofattributes? Inour context,the weightsare estimated by human judgement. However, the scale of relative