Ann.Telecommun. DOI10.1007/s12243-016-0550-7 Majority-consensus fusion approach for elderly IoT-based healthcare applications FaouziSebbak1·FaridBenhammadi1 Received:10January2016/Accepted:20October2016 ©InstitutMines-Te´le´comandSpringer-VerlagFrance2016 Abstract Nowadays, tremendous growth of Internet of Keywords Internetofthings·Healthcare·Evidence Things (IoT) applications is seen in smart environments theory·Combinationrule·Probabilitycalculus suchasmedicalremotecareapplicationswhicharecrucial reconcilement duetothegeneral agingofthepopulation.Withtherecent advancements in IoT-based healthcare technologies, activ- ityrecognitioncanbeusedasthekeypartoftheintelligent 1Introduction healthcare systems to monitor elderly people to live inde- pendentlyathomesandpromoteabettercare.Recently,the The rapid increase in the number of elderly people brings evidence theory and its derivates approaches began to take major issues to elderly health care including a rise in care placeinthefieldsofactivityrecognitioninthesesmartsys- cost,highdemandinlong-termcare,decreasecaregiverbur- tems.However,theseapproachesaregenerallyinconsistent den, and insufficient and ineffective care. With the recent with the probability calculus due to the lower and upper advancements in Internet of Things (IoT)-based healthcare probabilityboundsconsideringthecombinedevidences.To technologies, accelerometer, pressure, ECG, smartphone, overcomethesechallengesandtogetmorepreciselytherec- andassistedlivingtechnologies,activityrecognitioncanbe oncilementbetweentheevidencetheorywiththefrequentist used as the key part of the intelligent pervasive systems to approachofprobabilitycalculus,thisworkproposesanew monitorelderlypeopleforsustainingindependentlivingat methodology for combining beliefs, addressing some of homes and promote a better care [1, 17]. These intelligent the disadvantages exhibited by the evidence theory and its IoT-basedhealthcaresystemsarebecomingmoreprevalent derivatives. This methodology merges the non-normalized in every-day life by enabling the systems to adapt accord- conjunctive and the majority rules. The proposed rule is ingtotheelderlyhealthandpositioningcontext-awareness evaluated in numerical simulation case studies involving information. However, the contextual information cannot the activity recognition in a smart home environment. The be identifying precisely because of the limitations and the results show that this strategy produces intuitive results in imperfectionnatureofsensingtechnologiesduetothewear- favorofthemorecommittedhypothesis. able sensors’ inaccuracy. This imperfection of contextual informationwillbetranslatedtohigh-levelcontextsresult- inginbuildinginaccurateactivityrecognition.Agoodcon- textualinformationmodelingformalismreducestheimper- fectionofcontextualinformationandenhancestheaccuracy of activity recognition. So, one of the main challenges in (cid:2) FaouziSebbak [email protected] IoT-based healthcare applications is managing these forms of uncertainty due to sensor’s measurements which are FaridBenhammadi [email protected] pronetouncertainty.Anextensiveliteraturereviewinwear- ablesensor-basedactivityrecognitionanditsapplicationsin 1 EcoleMilitairePolytechnique(EMP),Algiers,Algeria healthcarehavebeencarriedoutoverthepastfewdecades. Ann.Telecommun. These applications are conducted on context representa- the context of health care and elderly care which repre- tion and reasoning in order to infer activity from accurate sent an important class of applications [35]. A major goal as well as inaccurate context in IoT-based healthcare solu- of current research in activity recognition in general is to tions [40]. Among the existing inference approaches, the enablehealthcareapplicationsfortheelderlypopulationto evidence theory [45] is a major constituent and commonly livemoreindependentlivesandreducetheburdenofcare- usedtohandleuncertaintyandincompletenessofcontextin givers.Forthisway,IoT-basedhealthcaretechnologiesand pervasive computing environments [12, 15, 19, 28, 39, 43, thefusionapproacheshelptoincreasethequalityandeffi- 67]. This theory offers a good alternative to some activity ciencyofcareofelderlypeopleandthusactivityrecognition recognition methods where domain knowledge is applied accuracy may be improved [22]. Villalba et al. [55] have [29] instead of training data that are not easily available in designed a system for vital elderly body signs which indi- ubiquitousenvironments.However,mostoftheapproaches cate imminent health threats in order to complement the based on the evidence theory produce incompatible results traditionalemergencyinpotentiallydangeroussituations.Si compared to the frequentist approach of probability cal- et al. [46] aim to support persons suffering from demen- culus. The main aim of this paper is to develop a novel tia through the use of context-aware reminders and similar methodformulti-sensor-basedactivityrecognitionofactiv- assistance.Wuetal.[61]havedevelopedasmartapproach ityofdailyliving(ADL)ofelderlypeopleforanintelligent thatclassifiescaneusageandwalkingpatterns,andinforms assisted living system in order to reach the reconcilement the elderly in case of high risk of falling. Barnachon et al. oftheevidencetheorywiththefrequentistappoachofprob- [5] present a novel framework for recognizing streamed ability calculus. Our approach is based on the merging of actionsusingmotioncapturedata.Theproposedframework twocombinationrules.Thismixtureallowstofavortheevi- aims at achieving early recognition of ongoing activities. dence with a big consensus because it is accordant to the Bravataetal.[51]proposedasystemwhichisabletoinfer objective of uncertainty reasoning over evidence accumu- elders’ health by monitoring the level of physical activity lation.Theproposedcombinationalternativeruleforactiv- performed. Based on assistive technologies for elders and ity recognition of elderly healthcare applications, called caregivers Ambient Assisted Living (AAL) systems, other Majority-ConsensuscombinationRule(MCR),isevaluated authorsarefocusedonimprovingcareprocessesbyprovid- and compared on Shafer’s model with the evidence the- ingelectronichealthrecords[21],remotemedicalattention ory and its most famous derivates. The results show that and suggestions [14], and intelligent vital signs monitors the present rule produces the intuitive results in favor of [30]. Monitoring and assessing the performance of daily the more committed activity and gives generally a high activitiesisrelevantforthecontinuousassessmentofelders’ consistency with the frequentist approach of probability health and independence [34]. The work proposed in [20] calculus. used the temperature sensor as part of their activity recog- Theoutlineforthispaperisorganizedasfollows:Section2 nitionsystems,e.g.,thedifferenceoftemperatureof15min presents the approaches used to carry out the research and is used to determine the use of a shower. The proposed the inference systems of the multi-sensor activity recogni- approachcanprovideadetailedactivitydetection;however, tionforelderlyhealthcaresystems.Section3introducesthe it suffers limitations in terms of feasibility. The approach backgroundofevidencetheoryanddecisionmakinginthis proposed in [11] allows to identify any changes in activi- theory.Section4describesthemathematicalfoundationsof tiessuchaschangeswalkingspeedandsleeprhythm.Some a new combination rule for uncertainty management strat- researchers have also introduced to older adults fall detec- egy.Theproposedcontext-awaresystemarchitectureofthe tion area using wearable device based, ambience sensor multi-sensor activity recognition and the simulated results based, and camera based [32]. Based on body acceleration are reported and discussed in Section 5. Finally, Section 6 sensors to measure the patients’ movements, the authors concludesthepaperwithfuturedirections. [3] provide a rhythmic auditory signal that stimulates the patienttoresumewalking.Hossain[16]introducedthecon- cept of virtual caregivers to assist the caregiver in making 2Reviewofsensor-basedactivityrecognition somebasicdecisionsforhealth-relatedapplicationstomon- approaches itor elderly people to live more independent daily lives. The proposed system uses context information to promote In the past decade, a variety of sensors have been used moreactiveandthushealthylifestyle,ortoactivelysupport to allow the elderly patients with diseases to achieve con- elderlyordisabledpeopleinperformingeverydayactivities. tinuous monitoring of their health condition [1, 2, 6, 31]. Analyzing more heterogeneous sensor sources, the system Currentresearchinactivityrecognitionfromwearablesen- make elders’ health inferences. Chernbumroong et al. [8] sors covers a wide range of topics, with research groups proposed a novel multi-sensor-based activity recognition focusing on topics such as the recognition of ADLs in system.Theapproachusesmultiplelow-cost,non-intrusive, Ann.Telecommun. non-visualwearablesensorsonthewrist.Thesensorfusion multipleindependentsourcesdefinedwithinthesameframe is performed at two levels (feature and classifier level). of discernment. Based on Shafer’s model of the frame (cid:2), Other works are focused on a predictive schedule system Dempster’s rule is defined by the following equations for thatblendscaregiverusingsensormeasurementsinsteadof twosources: trying to substitute the caregiver with an automatic caring mc (X) m (X)= 12 (2) system [52]. To infer situations and activities in the con- DS 1−mc (∅) 12 text of elderly health care, the preceding applications use (cid:2) Bayesian networks [12, 39, 64], dynamic Bayesian net- mc (X)= m (X )m (X ) (3) 12 1 1 2 2 work(DBN)[38],conditionalrandomfield(CRF)[19,54], X1,X2∈2(cid:2) and hidden Markov models [9, 19, 36, 59]. These meth- X1(cid:2)∩X2=X odsareusuallybasedontrainingdatatorecognizeactivities mc (∅)= m (X )m (X ) (4) 12 1 1 2 2 inferredfromlower-levelsensorinformation.VanKasteren X1,X2∈2(cid:2) et al. [19] use hidden Markov models to infer high-level X1∩X2=∅ activitiesofaperson’sinthehome.Intheproposedmodel, where mc (X) and mc (∅) represent the conventional con- 12 12 activitypatternsovertimearelearnedfromtrainingdata.In junctiveconsensusoperatorandconflictofthecombination [13],theauthorsproposetheuseofhiddenMarkovmodels betweenthetwosources,respectively. to classify six different activities. Bayesian networks were From a given bba m, the decision functions are defined used to determine the activities from lower-level sensor asfollows: data. Wang et al. [58] investigate the problem of recogniz- - The belief function Bel(X) measures the belief that ing multi-user activities using wearable sensors in a home hypothesisXistrueanditisgivenby: (cid:2) setting.Theauthorsdevelopamulti-modal,wearablesensor Bel(X)= m(Y) (5) platformtocollectsensordataformultipleusersusingtwo Y⊆X temporal probabilistic models, the coupled hidden Markov -TheplausibilityfunctionPl(X)canbeinterpretedasa model(CHMM)andthefactorial conditionalrandomfield measureofthetotalbeliefthathypothesisXcanbetrueand (FCRF).In[25],theauthorsusethehierarchicalconditional itisgivenbythefollowingformula: randomfieldstoderivehigh-levelactivitiesandtoidentify (cid:2) placesfromGPSdata.Recently,theevidencetheorybegan Pl(X)= m(Y) (6) to take place in the fields of activity recognition in smart X∩Y(cid:6)=∅ environments and ubiquitous computing [15, 28, 60, 67]. -Thepignisticprobabilitytransformation[48,50]isgen- Wu [60] uses the DS theory to fuse the sensor evidence in erallyconsideredasagoodcriteriaforadecisionrule.Itis hissensorfusionmodel.Wuusesastaticweightingonsen- definedforallX ∈2(cid:2),withX (cid:6)=∅;by: sormassfunctionstoindicateevidencereliability.Tomodel (cid:2) |X∩Y| m(Y) the activity structure, a situation directed acyclic graphs BetP(X)= (7) (DAGs)isproposedin[28]andtwotypesofevidentialnet- |Y| 1−m(∅) Y∈2(cid:2) works: activities-activity and sensors-contexts-activity are Y(cid:6)=∅ proposedin[15,43]. However, some researchers argue that this combina- tion rule has an abnormal behavior when the conflict between sources becomes high [41, 42, 47, 56, 66]. So, 3Theevidencetheoryanditsderivates several attempts have been proposed to avoid this abnor- malbehavior.TheseattemptstomodifyDempster’srulecan The evidence theory [45] has attractive properties which be divided into two main categories : corrective-evidence provide significantly richer information in the fusion field. approaches [7, 18, 23, 24, 33, 62, 65] and the conflict BasedonShafer’smodel,theframeofdiscernmentisaset redistribution approaches [10, 26, 47, 50, 63]. The first of mutually exclusive and exhaustive hypotheses about the approachesconsistofcorrectivestrategiesoftheinitialbasic problem domains. From a frame of discernment (cid:2) corre- belief assignments while using thereafter Dempsters rule spondingly, 2(cid:2)is the power set of (cid:2), then a basic belief for combining these corrective evidences. The idea behind assignment (bba) or proper mass is defined as a mapping the second approaches is to transfer conflicting masses m(·): proportionally to empty or non-empty sets according to (cid:2) somecombinationresults.However,theproblemwiththese m(∅)=0and m(X)=1 (1) approachescanberelievedtosomeextentbyreplacingthe X∈2(cid:2) evidenceswithinappropriatelysmall-weightsevidenceval- The Dempster’s combination rule is the normalized con- ues. Moreover, all these approaches are inconsistent with junctive operation which aims to aggregate evidence from theprobabilitycalculuswheretheinformationcombination Ann.Telecommun. results are sometimes counterintuitive and very far com- opinion, and unearned certainty. However, using original pared to those obtained by the probability calculus. This evidence sources and the results of the conjunctive con- abnormal behavior and the inconsistency of these derivate sensus, PCR6 rule redistributes an evidential conflict in rules will be illustrated through some examples studied in proportiontothefocalsetelements,whichisacompromise thefollowingsections. strategyandgivesmoreintuitiveresults. Table 1 reports the combination results and comparison For the disappearing ignorance, Pearl criticized the dis- ofexistingcombinationrulesusingsomewell-knownexam- appearing ignorance [37] caused by the operation of inter- plesdiscussedintheliteratureswhereDempster’sruleand sectioninDempster’srule,andMurphyproposedacombi- some alternative combination rules give counter-intuitive nationwhichdoesnothavetheeffect[33].Forthecertainty results.Theunderlinecombinationresultsinthetableshow convergence case, Yager, Smets, and Dubois rules fail to that the results are judged to be counter-intuitive. As it convergetowardthefocalelementsupportedbybothpieces can be seen from Table 1, when the conflicting mass is ofevidence.Theproblemoftotalcertaintytominorityopin- computed, the Dempster’s rule for combining evidences ioncausedbytheoperationofintersectionintheDempster produces counter-intuitive results that do not reflect the rule[66]issolvedinmanyalternatives. actualdistributionofbeliefsinsome cases. Thislaterdoes TheoperationofintersectionintheDempsterrulecauses not reflect the actual distribution of beliefs in some cases also the exclusion of elements which are not contained in suchaslossofmajorityopinion,totalcertaintytominority focal elements of one or more bodies of evidence. Known Table1 Comparativeresultsforwell-known Rules Cases Dempster Murphy PCR6 Yager Smets Dubois Disappearingignorance m1({a})=0.5 m{a}=0.5 m{a}=0.3571 m{a}=0.5 m{a}=0.5 m{a}=0.5 m{a,b,c}=1 m1({b})=0.5 m{b}=0.5 m{b}=0.3571 m{b}=0.5 m{b}=0.5 m{b}=0.5 m2({a,b})=1 m{a,b}=0.2857 Certaintyconvergence m1({a})=0.5 m{a}=0.6667 m{a}=0.6667 m{a}=0.625 m{a}=0.5 m{a}=0.5 m{a}=0.25 m1({b})=0.5 m{b}=0.3333 m{b}=0.25 m{b}=0.375 m{b}=0.25 m{b}=0.25 m{a,b,c}=0.75 m2({a})=0.5 m{a,b}=0.0833 m{a,b}=0.25 m{∅}=0.25 m2({a,b})=0.5 Totalcertaintytominorityopinion m1({a})=0.9 m{b}=1 m{a}=0.488 m{a}=0.486 m{b}=0.01 m{a}=0.01 m{b}=0.01 m1({b})=0.1 m{b}=0.024 m{b}=0.028 m{a,b,c}=0.99 m{∅}=0.99 m{a,b}=0.09 m2({b})=0.1 m{c}=0.488 m{c}=0.486 m{a,c}=0.81 m2({c})=0.9 m{b,c}=0.09 Lossofmajorityopinion m1({a})=0.9 m{c}=1 m{a}=0.7336 m{a}=0.3958 m{c}=0.01 m{a}=0.98 m{a,b}=0.36 m1({a,c})=0.1 m{b}=0.1657 m{b}=0.1305 m{a,b,c}=0.99 m{a,c}=0.02 m{c}=0.01 m2({a,b})=0.8 m{c}=0.0477 m{c}=0.1342 m{a,b,c}=0.54 m2({a,c})=0.2 m{a,b}=0.0503 m{a,b}=0.307 m3({b})=0.5 m{a,c}=0.0027 m{a,c}=0.032 m3({c})=0.5 Unearnedcertainty m1({a})=0.5 m{a}=1/3 m{a}=0.3 m{a}=0.375 m{a}=0.25 m{a}=0.25 m{a,b}=0.25 m1({b,c})=0.5 m{b}=1/3 m{b}=0.2 m{b}=0.25 m{b}=0.25 m{b}=0.25 m{a,c}=0.25 m2({c})=0.5 m{c}=1/3 m{c}=0.3 m{c}=0.375 m{c}=0.25 m{c}=0.25 m{b,c}=0.25 m2({a,b})=0.5 m{a,b}=0.1 m{a,b,c}=0.25 m{∅}=0.25 m{a,b,c}=0.25 m{b,c}=0.1 Ann.Telecommun. asthelossofmajorityopinion,thiscaseissolvedbyafew where m (X) and m (X) denote the majority and the Maj Conj alternativecombinationrules[27,33,63]. non-normalized conjunctive rules of the focal element X, The problem of unearned certainty known as unwar- respectively. ranted mass is also caused by the operation of intersection The non-normalized conjunctive rule, given by Smets intherule.Itissolvedbymostofthealternatives. [49]isdefinedasfollows: (cid:2) (cid:5)s (cid:3) (cid:4) m (X)= m X (9) Conj i j 4Proposeduncertaintymanagementstrategy X1,X2···,Xj∈2(cid:2)i=1 X1∩X2∩···Xj=X The evidential reasoning is an important technique of uncertaintycontextmodelinginmulti-sensor-basedactivity Themajorityruleisdefinedby: recognition in elderly healthcare applications. The present (cid:2)s sectionintroducesourcombinationmergingruletechnique m (X)= m (X) (10) Maj i tomanagetheuncertaintyforactivityrecognitioninhealth i=1 careapplications.Thestartingpointofourcombinationrule A general measure of the global conflict in evidence is investigations is to describe the strategy for combining the definedbythefollowingformula: belief functions. Thereafter, we introduce a new notion to verify the consistency of the combination rule of evidence (cid:2) (cid:5)s (cid:3) (cid:4) withtheprobabilitycalculus.Inthiswork,wemadeastrong k = m X (11) i j assumption that security is guaranteed by using an appro- X1,X2···,Xj∈2(cid:2)i=1 priatemechanismsuchastheuseofacryptographicsystem X1∩X2∩···Xj=∅ ensuringanend-to-endsecurity. where X ∈ 2(cid:2) denotes a response of the information j sourcei,andm (X )denotestheassociatedbelieffunction. 4.1Theproposedruleprinciple i j This formulation of the combined mass yields the next result. The evidence theory and its derivates have given birth to a large family of rules to combine multiple evidences. To overcome the drawbacks of the previously presented rules, Theorem1 Themassm1,2,···,s|MCR(·)isapropermass. we propose an alternative combination rule which mixes the non-normalized conjunctive rule, given by Smets [49] Proof A proper mass means that a massm 1,2,···,s|MCR(·) (theconventionalconjunctiveconsensusoperatorintheevi- must satisfy all requirement given by Eq. 1. WhenX is an dence theory) and the simple fusion rule based on the emptyset,thevalueofm(X)iseq(cid:3)ualto0.Bydefinition,(cid:4)the majorityrule.Theideabehindourcombinationruleisthat massm1,2,···,s|MCR(X)= s+11−k mMaj(X)+mConj(X) . it redistributes the global conflict into already implicated Hence, (cid:6) focal element setsenteringin themajority andtheconsen- (cid:2) (cid:2) (cid:3) 1 susresults. Thisruleallowsto bringabout areconciliation m1,2,···,s|MCR(X) = s+1−k mMaj(X) between the belief functions theory and the probability X∈2(cid:2) X∈2(cid:2) (cid:7) calculus. (cid:4) +m (X) The formulation of our rule is based on the normaliza- Conj ⎡ tion of the sum of masses fused with the result of masses (cid:2) obtainedbythemajorityruleandthoseobtainedbythecon- = 1 ⎣ m (X) junctive consensus operator. So, our rule transfers the total s+1−k Maj X∈2(cid:2) ⎤ conflicting mass and total masses to non-empty sets pro- (cid:2) portionally to their resulting masses obtained by these two + m (X)⎦ operators. Conj Theproposedcombinationruleformultiplesources(s ≥ X∈2(cid:2) 2)isformulatedasfollows: Bydefinitionofthemajorityandconjunctiverules,itiseasy (∀X (cid:6)=∅)∈2(cid:2) toseethat: (cid:2) 1 (cid:3) (cid:4) m (X)=s m1,2···,s|MCR(X)= s+1−k mMaj(X)+mConj(X) (8) X∈2(cid:2) Maj Ann.Telecommun. and heart rate monitor. Assume that the resulting training (cid:2) m (X)=1−k. data induces the following masses bba’s for activity Conj recognition. X∈2(cid:2) Then, we get ss++11−−kk = 1. Therefore, m1,2,···,s|MCR(·) is a m1(Falling)=0.8 m1({Falling,WatchingTV})=0.2 propermass. m (Falling)=0.1 m (Sleeping)=0.9 2 2 As an example, consider a simple monitoring for an m (Falling)=0.25 m ({Exercising,WatchingTV})=0.75 3 3 emergency detection system. For illustration, we treat a scenario (see Fig. 1) of health monitoring example for an According to these masses, the classes’ distribution for elderly people activities of daily living. In order to shift each activity are : the accelerometer indicates that the the care to a personalized level, we consider two kinds of elderly people is falled down with 80% and the remain- InternetofThingssensors:thosethatuseexternalwearable ing 20% are imprecise recognition rate for two activities sensors that are attached to different parts of the elderly {falling} or {WatchingTV}. The smartphone indicates body and those that use embedded inertial sensors of the that the elderly people sleeping with 90% while 10% smartphonewhereitisnotmandatorytomountorfixposi- reflectsthathe/sheisfalling.However,theheartratemon- tion of the sensor. Our scenario consider four activities: itor indicates that imprecise recognition of 75% of the exercising,watchingTV,sleeping,andfalldetectionwhich elderlypeopleexercisesonanabdominalbenchorwatching is another important area of emergency detection, and can TVand25%forelderlyfalldetection. be especially useful for the elderly. Figure 1 visualizes a Weassumethateachsensorprovidesitsactivityselection scenario in which an elderly patient’s health profile and independently. The non-normalized conjunctive consensus vitals are captured using portable medical devices (wire- of the basic belief assignments yields the following results less accelerometer and heart monitor rate) attached to his forthisexample: or her body and smartphone for communication. Captured datafromthesemedicaldevices arethencommunicated to mConj(Falling)=0.025 mConj({Falling,WatchingTV})=0 the smartphone and thereafter, stored data become useful mConj(Sleeping)=0 mConj({Exercising,WatchingTV})=0 for activity recognition, aggregation, and inference. Based on aggregation and evidential inference, doctors or care- The global conflicting mass k is 0.975 and the masses 123 giverscanmonitoranelderlypatientfromanylocationand fusedwiththemajorityruleare: respond accordingly in risk situation such as elderly fall detection. Supposing that a dataset contains data from these sen- mMaj(Falling)=1.15 mMaj({Falling,WatchingTV})=0.2 sors:wirelessaccelerationsensor,smartphone,andwireless mMaj(Sleeping)=0.9 mMaj({Exercising,WatchingTV})=0.75 Fig.1 Scenarioillustrationfor elderlymedicalcare Ann.Telecommun. ByapplyingEq.8,thefinalcombinedmassesare: m123|MCR(Falling)=0.388 m123|MCR({Falling,WatchingTV})=0.066 m123|MCR(Sleeping)=0.298 m123|MCR({Exercising,WatchingTV})=0.248 Using the maximum of the plausibility or the pignis- Accordingtobeliefandplausibilityfunctionsandthefre- tic probability to provide a decision from the remaining quentistapproachofprobabilitycalculus,thereallowerand combinedmasses,weselectthefallingsituationthatoccur upperboundsofanunknownprobabilitymeasureisdefined (Pl(Falling) = 0.455andBetP(Falling) = 0.421)asa on(cid:2)andcompatiblewiththebasicbeliefassignmentm(·) plausibleactivitythattheelderlypeopleisdoing.Thus,the bythefollowingdefinitions. smartphone send an alert signal alarm to the hospital and notifiedandprovidedanavailablecaregiverwithinventory Definition 1 (Real lower bound ) Let (cid:2) be a frame of information of elderly activity. In this case, it is necessary discernment and let mi(·) be the basic belief assignments toprovideanalertsignalforavailablecaregiverandambu- associatedtoagivenfocalsetXi ∈ 2(cid:2).Wedefinethereal lance driver within seconds of a risky situation to prevent lowerboundofthesingletonfocalelementX ∈(cid:2)(denoted a compromise to an elder’s health. This result involves a BelL(X))asfollows: continuous supervision which is due to the fact that most (cid:2)s 1 of basic belief assignments are in favor to this situation in BelL(X)= m (X) (12) i respectwiththesensorsdatacollection.Asaresult,ourrule s i=1 providesagoodevidencecombinationfordatafusioninthe waythatitpreservesthemajorityandtheconsensusofeach Definition 2 (Real upper bound) Let (cid:2) be a frame of sensor. discernment and let m (·) be the basic belief assignments i The present rule is commutative since the conjunctive associatedtoagivenfocalsetsX ∈2(cid:2).Wedefinethereal i and the sum operators have this property. However, it is upperboundofthesingletonfocalelementX ∈(cid:2)(denoted not associative as the most evidence theory derivates but PlU(X))asfollows: can become quasi-associative if we save the result of the (cid:2) 1 conjunctive and majority rules at each step before the PlU(X)= m (X ) (13) i i s normalizationprocess.Unfortunately,ourruledoesnotpre- Xi∩X(cid:6)=∅ serve the vacuous belief assignment when some sources become totally ignorant. The disappearing ignorance pro- Thereallowerbounddefinitioncorrespondstotheaver- prietyisanadvantageincombinationprocessbecauseonly age of evidence for each hypothesis X. In other words, it the evidences without ignorance are combined. However, represents the fusion using the arithmetic mean operator. some authors [33, 37, 56, 57] have criticized this propri- However, the real upper bound measures the average mass ety caused by the conjunctive operation and specified that of the total mass which can visit somewhere within X but thecounterintuitiveresultsofDempster’srulearecausedby maymoveoutsideaswell(eachmassoftheignorancecon- thisignorance.Hence,inourapproach,wecannotsuppress taining X is added to the mass of the focal element {X}). or neglect the weaker belief committed to the area’s activ- Obviously, we have BelL(X) ≤ PlU(X) for all X ∈ (cid:2). ityofelderlypeoplewhichcanbesupportedbythevacuous Notethattheserealboundsaredifferentfromthebeliefand beliefassignmentbecausewepresumethatthedisappearing plausibilityfunctionsdefinedintheDempster-Shafertheory ignoranceisnotjustifiedinallsituations. (seeSection3). These two definitions represent the reconcilement inter- 4.2Combinationrulereconcilementconsistency val of a combination rule of evidences with the frequentist approachofprobabilitycalculus. The notion of the combination consistency represents the Let us reconsider our elderly people example. To show interval of reconcilement (approximation) of the combina- theinconsistencyofsomewell-knownruleswiththeprob- tionofevidenceresultcomputedbythefrequentistapproach abilitycalculus,letusexaminetheactivityselectionsinthe of probability calculus and the evidence theory and its reality.Hence,wecomputetherealinterval(thereallower derivates. and upper bounds) of each activity selection when every Ann.Telecommun. Table2 Thereallowerandupperbounds TVandexercisingactivities([0;0])couldneverhavebeen selected according to the real probability of the selection Elderlypeopleactivity Lowerbound Upperbound whicharegivenbythethreesensors.ThePCR6ruleselects Falling 0.383 0.450 the abnormal activity detection, sleeping with the pignis- Sleeping 0.3 0.3 tic probability equals to 0.377 which is outside the real Exercising 0 0.250 lower and upper bounds interval ([0.3;0.3]). This abnor- WatchingTV 0 0.317 mal behavior of this rule is due to the absorption of the conflict by the focal element sleeping activity which has a bigger mass (m (Sleeping) = 0.9) in the conflict redistri- 2 butionstep.Murphy’srulefavorsthefallingsituationbythe sensor has finally selected only one activity at the same averaging technique of its masses and gives the lower and time. For this example, the real lower and upper bounds upperboundsofthefallingsituationequalto[0.607;0.609]. canbedefinedsupposingthethreesensorsaregatheredinto Thisresultisoutsidetherealbounds([0.383;0.450]).How- a single sensor. Hence, applying equations 12 and 13, we ever, the result obtained by our rule are reasonable for have obtained the results reported in Table 2. In the first this example since the belief, the plausibility, and the pig- row,thereallowerboundmeansthatonlythepreciseselec- nistic probability of each activity’s selection is inside the tions of the falling situation (activity) are considered by interval of the real bounds except for the sleeping activity the three sensors. However, the real upper bound of this whichapproximatestherealprobability(BetP(Sleeping)= activityassumesinadditiontothepreciseselectionthatthe 0.298). accelerometer sensor had finally selected the falling situa- tion. Clearly from these real bounds, the final selection by thethreesensorscorrespondstothattheelderlypeoplemust 5Simulationstudies be an emergency since the real lower bound of the falling situationisgreaterthaneachrealupperboundoftheother The foregoing sections have provided background on the activities. proposedcombinationrule.Now,wereturntoevaluateour The results reported in Table 3 show the inconsistency rule with simulation studies based on the elderly health- of the evidence theory and its derivates to compute the care scenario cases. The first part of this section describes lower and upper bounds compared to the results reported our uncertain context information modeling in IoT-based in Table 2. Clearly based on Dempster-Shafer rule, the healthcaresystems.Thesecondpartdescribesexperimental lower and upper bounds specify that sleeping, watching simulationsaccordingtoelderlyhealthcarescenariocases. 5.1Uncertaincontextmodeling Table3 Thelowerandupperboundsaccordingtotheprecedingrule classes IoT-based healthcare systems for elderly people need to represent many different types of uncertain context infor- Elderlypeopleactivity Rule [Bel(X);Pl(X)] BetP(X) mation, such as sensor information, spatial information, Falling Dempster’s [1;1] 1 medical history, and activities profiles. As introduced in Murphy’s [0.607;0.609] 0.608 [44], the context can be divided into three layers or levels: PCR6 [0.3407;0.365] 0.353 sensors, abstracted context, and situation (or activity). The Ourrule [0.388;0.455] 0.421 sensor layer refers to the physical layer where sensors or Sleeping Dempster’s [0;0] 0 anyotherIoTdevicesareusedtoprovidecontextualinfor- Murphy’s [0.181;0.181] 0.181 mation such as temperature, acceleration and motions, etc. PCR6 [0.377;0.377] 0.377 Inthesecondlayer,rawdataprovidedbythefirstlayerare Ourrule [0.298;0.298] 0.298 translated into more human understandable representation Exercising Dempster’s [0;0] 0 by using abstraction. For example, accelerometer reading Murphy’s [0;0.104] 0.052 are translated to sates of “in motion,” “falling,” or “immo- bile.”Thishumanunderstandablerepresentationofcontext PCR6 [0;0.258] 0.129 will follow a formal representation in order to be used in Ourrule [0;0.248] 0.124 pervasivesystems such as theontologymodel. Thismodel WatchingTV Dempster’s [0;0] 0 is particularly interesting to provide an explicit commonly Murphy’s [0.106;0.212] 0.159 agreeduponrepresentationofconceptsinahierarchalman- PCR6 [0;0.282] 0.141 ner. The last layer is the high-level context abstraction Ourrule [0;0.314] 0.157 defined by fusing multiple information contexts to infer Ann.Telecommun. activity or situation. Then, this infrastructure can employ and home environmental sensors are deployed to gather as the IoT health care to enable communication between much information about and around the elderly as possi- elderlyindividuals,caregivers,anddoctors.Inthenext,we ble.Thelowerlayerrepresentsasetofcontextsensorsand explain our uncertain context representation and reasoning actuators which encapsulate the hardware devices of IoT- mechanisms. based healthcare technologies. The second layer is called Our ubiquitous computing framework proposed in [44] the OSGi layer. The goal of this layer is to acquire con- aimstobuildnewapplicationstoprovideassistiveservices text information from the ubiquitous environment and to for people in autonomously simulation way. These frame- control devices and equipments. The third layer represents work must adapt their behaviors to provide services that the multi-agent layer. This layer aims to interpret, infer, match current user needs. The proposed approach presents and share uncertain contextual information data. Thus, we acontext-awareubiquitousframeworkbasedonlightweight use the ontology markup language OWL for our ontology coupling between multi-agent system and the OSGi (Open context-based representation to share information acquired ServiceGatewayinitiative)framework[53].Figure2illus- from the environment or aggregated by the agents. In our tratesthearchitectureofourubiquitousframeworksystem. ubiquitous framework, we have extended USARSim sim- Inthesmarthomehealthcarescenario,variouselderlybody ulator [4] as simulation infrastructure for the proposed Fig.2 Overviewofthemulti-agent–OSGIcontext-awareframework Ann.Telecommun. elderly’s IoT-based healthcare system. Some simulation andriskysituations(exercising,sleeping,falldetection,...), images from simulated configurable sensors and environ- and (iii) notification of emergency situations indicating a mentsareshowedinFig.3. health risk by sending signal alarm to notify hospital staff In our ubiquitous framework system, the evidential rea- (doctor,caregiver,ambulancedriver,...).Forexample,when soningagentiscapableofperformingthefollowingreason- smart home monitor agent localized current elderly people ing tasks: (i) continuous information contextualization of localization and evidential agent infers fall detection situa- thephysicalstateofanelderlyperson(motion,localization, tion, it sends to the emergency agent the following request medical monitors,...), (ii) recognition of possibly activities medicalassociatedrule: Rule: IF (ElderlyHasFallenDown)AND(HeartRateisVeryHigh)AND (RespirationRateisLow)AND(ElderlyDoNotRespondedToSign(phonecall)) Then (SendEmergencySignaltoHospital)AND(ProvideElderlyInventory InformationtoanAvailableCaregiver)AND(CallAmbulanceDriver) Thisrequestruledescribesthatiftheelderlypersonhas Using these masses, one gets the following lower and fallen down and after he/she has been notified but he/she upperbounds has not responded yet, then the system send an emergency signalalerttothehospitalandanavailablecaregiverreceive (cid:2)3 1 anelderlypersoninventoryinformationandtheambulance BelL(Falling)=PLU(Falling)= a i 3 driveriscalledtoo. i=1 (cid:2)3 1 5.2Experimentalsimulationcases BelL(Sleeping)=PLU(Sleeping)= bi 3 i=1 5.2.1TheBayesianbbasselections BelL(WatchingTV)=PLU(WatchingTV) (cid:2)3 1 The experimental simulations are designed and imple- = (1−a −b ) i i 3 mented based on a modified USARSim simulator [4] and i=1 extended multi-agent OSGi inference mechanism intro- duced in [44]. In this section, we are going to present To illustrate the consistence of the combination rules, the behavior of our rule to combine bodies of evidence in wehaveimplementedaMonte-Carlosimulationwherethe Shafer’s model using the Bayesian bbas. Suppose that the masses were distributed randomly according to a uniform sensors mentioned in our simplest elderly monitoring for distribution.Thus,wehavegeneratedrandomly1000sam- emergencydetectionwhereeachsensorinfersonepreferred ples for the three sensor selections. Figure 4 illustrates the activityorsituationasfollow: consistency results of our rule compared to the evidence Accelerometersensor:Falling<Sleeping<WatchingTV theory and its derivates for the elderly fall detection. For a reason of clarity of the presented comparison results, m1(Falling)=a1 m1(Sleeping)=b1 we show only the results for 100 samples. The frequen- m (WatchingTV)=1−a −b tist approach of probability calculus was initially used 1 1 1 to calculate the lower and upper bounds with respect to Smartphone:Sleeping<WatchingTV<Falling activity selections used here. We report the relative com- m (Falling)=a m (Sleeping)=b parisonoftheselectionresultforourruleandtheevidence 2 2 2 2 theory and its derivates (Murphy’s and PCR6 rules) based m (WatchingTV)=1−a −b 2 2 2 on the maximum pignistic of probabilities criterion. It is Heartmonitorsensor:WatchingTV<Falling<Sleeping clear from this figure that, in the Bayesian case, our rule has a high approximation of the real probabilities where m (Falling)=a m (Sleeping)=b 3 3 3 3 the majority of the pignistic probability values are near m (WatchingTV)=1−a −b the real probability values computed from the frequentist 3 3 3