Data Mining in Medical and Biological Researchnd and cost US$487, 000(Nohzawa, 2003) The latest CPU, the Yorkfield XE, contains 820lion transistors on 2 x 107 mm2 dies and features a 1333 MT/s FSB and a clock speed of 3Since the first integrated circuit (IC), which contained a single transistor and severaled by Jack Kilbylogy makes it possible to condense an entirecomplicated spread-spectrumnication system into a magic box as small asinterconnected through wired and wnetworks, andnvisibnment This makes " informationlight speed"possible We are already acclimatized to enjoy everythingorldwide conveniently, wherever we are We enjoy online shopping and share informationwith friends from the other side of the earth in an instantHowever, thisble-edged sword Our daily lifestyle has changed dramatically whilwe may benefit from the advantages of today s society, at the same time, we face manyunprecedented problems in the health domain, which have emerged with all of thesechronic illnesshas occurred concurrentlywith the accompanying lifestyle changes
Figure 1 shows the change in mortality amddifferent diseaseshe past 100Japan There has not beenarge change in conventional causes of death, such as contingency, caducity, andpneumonia Acute infectious diseases, such as tuberculosis, have disappeared completelysince the 1980s Howeath due to chronic conditions is increasing Theneoplasm), which account for 60 per cent of total deaths00m0000000othersdiseasecardio diseasemalignant neoplasm聖至och(year)Fig 1 Change in mortality of different diseases over the previous century in Japan(Adapted from the Japanese Ministry of Health, Labour, and Welfare
A Scalable Healthcare Integrated Platform( SHIP)and Key Technologies for Daily Application0+個個1546ovement period in seconds (a) Measured pressure signal distorted by body movementsb Reconstructed well as the detected boriod (c) Reconstructed waveform and detected BR as well as the detected bodIIL LiLLI■I■ⅢIIlLFig
6 Two profiles of the BR/HR obtained from measurements over a single night Thesolid dots and vertical bars, terminated at thelines, show the mean values and standard deviation within a period of one minute Thedy movement periods are indicated by the variable-width vertical ba
Data Mining in Medical and Biological Researchhouse over a period of seven months, under an informed agreement for the use of the datafor research purposes During this period, data over about 30Therefore, data from a period of 180 d were recorded The data are plotted on a day-by-daybasis The verticarepresents the BR/HR in units of bpm The symbols and verticalvalues and standard deviation of thenight
The bold line is derived by filtering the mean values of the HRndow The dashed line isday base heart ratet is observed that the profile of the meanresponds to the female monthly menstrual cy040304/2305/1306/0206/2207/1208/0108/2109/1009/3010/20Time, dayday basis The symbols andof the BR/HR over 180 nights Data are plotted on a day-by-horizontal lines, show the mean values and standard deviation of the detected HR(o) andhe Br()in thThe devised system is completely invisible to theneasurements "Plug is allof its significant characteristics in securing perpetuo is just to plug in an AC power cable and a Lan cableven forget the exisR用mpressure under the pillow is detected as a first-order signal Theand body movements are derived as second-order parameters Sleep stage
A Scalable Healthcare Integrated Platform(SHIP) and Key Technologies for Daily Application 189arameters In the end, a comprehensive interpretatarallab studies, and also for screening patient property would increase its applicability in sleepbtained and a fully automatic operationdo not need a full sleep diagnosisly3 Wearable monitor for bodmultiHowever, a reliable evaluatircle based on the Bbt requient of ature under constantconditions for long periods
It is indeed a tedious task for a woman to measure her oral orarmpit temperature under similar conditions when she wakes up every morning over a longnin under theer,the traditional method for evaluating ovulationnatrual cycle dynamics iniricalts of BBT It has been pointed out that the BBt failed todemonstrate ovulation in approximately 20% of ovulation cycles among 30 normallylity and the accuracy of thand a Hidden Markov Model (HMM) based a statisticalroach to estimate the biphasic properties of bodhe menstrualthe cutaneous temperature around the abdominalbetween the breasts工+ perature senneQR-eodeLFig 8 A small, lightreight =59 g) for cutaneoussleep(QOL Co Ltd, 2008)
Data Mining in Medical and BiologicalThe device ismeasure temperature over 10 min intervals from midnight to 6am At most, 37 data points can be collected during the six hours Outliers above 40C orcode, known as"Quick Response"code(QR code)(Denso Wave Inc, 2000)and depictedLCD display As shown in Fig 9, the user uses the camera built into a mobile phone todnd transmitted to a databaobile network for data storage and physiological interpretation through data mining題Fig 9 A procedure for temperature data collection using a wearabphone(a)A QR code image(b)A QR code captured by a camera built into a mobile phone
) Original data recovered from the image captured by a mobile phone(QoL Co LtdThe temperature data measured during slee10(a) The nightly data are plotted in the vertical direction and have a range of 32 to 40The purpose of data mining in this study wasmperature profile during the menstrual cycle freAs shown in Fig 11, the biphasic properties of the menstrual cyclebe modelleddiscrete Hidden Markov Model (HMM)with two hidden phases The measuredemperature data are considered to be observations being generated by the Markov processing to the probability distribution The probability bl()is indicative that thevalue k is generated from the hidden LT phase The probability bH(k)is indicative that thevalue knerated from the hidden HT phase The probabilityse transition between LT and hT phasesFigure 10(b)shows the results after pre-processing to remoutliers from the raw datand eliminating any discontinuities from non-data-collectionFigure 10(c) shows theHMM estimation output using the pre-processed data fromb) as the input Figureo(d shows the estimafteruperimposed black symbols"*denote the menstrual periods recorded by the ustransition from the hf phase to the Lt phase denotes a menstrual period, while the reversetransition denotes ovulation
A Scalable Healthcare Integrated Platform( SHIP)and Key Technologies for Daily Application2910/19110811/2812/1801/0701/2702/1603/0803/2804/1750010001500200025003000350040004500Data No5002000300035004000LAnnenture dataeasured over a period of six months (b) Pre-processed results (c) Biphasic estimationroach (d) Post-pid results The symlmenstrual period recorded by the user
bLCkphaseFig 11 A discrete hidden Markov model with two hidden phases for estimating biphasicproperty in a menstrual cycle from cutaneous temperatureThe biphaseFig, 10(c)estimated by finding an optimal HMMmeter set that determines the hiddeon a given series of measured temperature data, as shorFig 10(b) The parameter set(A, B, )is assigned randomly in the initial condition and optimized through the forwardbackward iterative procedure until P(O 2) converges to a stableuntilhe absolute logarithm of the previous and current difference in P(o )is not greater than 8
Data Mining in Medical and Biological ResearchThe algorithm for calculating the forwthe backward variable, B, and theforward-backward variable, y, are shotat time, tobservation sequence, O1,O,, O until time f, and can be calculated using the following)=P(O1,O2Oa1()=rb(O)1≤i≤N,t=12a,(a, b,(0The backward variable, B (), denotes the probability of phase, qi at time, t, based on a partialbservation sequence, O+1, O +2,--, Or, from time t+l to T, and can be calculated using thefollowing steps for a given set of A(A, B,l=4()=1,1≤To find the optimal sequence of hidden phases for a given observation sequgiven model, M(A, B, r), theretiple possible optimality criteriaChoosing the phases, qu that are individually most likely at each time,P(qt-ilO, ) is equivalent to finding the single best phase sequence (path), ie
, maximizinQlO, )or P(Q,OA) The forward-backward algorithm is applied to find the optimalsequence of phases, gu at each timeobservation sePIO(233)i)(1)∑a(i)B()The most likely phase, gu
A Scalable Healthcare Integrated Platform( SHIP)and Key Technologies for Daily Application 19l≤t≤T2-3-4)As thereexisting analytical methods for optimizing A(A, B, ) P(o A)or P(O, 1|A)isally maximized (ie,A=arg max P(oIA) or i=argmax p(o, Q1a)) using gradientchniques and an expectation-maximization method
In this study, the Baum-Welchhe Baum-Welch re-estimation algorithm, we defined a variableEt(iD), to express the probability of a datum being in phase i at time t and phase jat time t+given the model and the observation sequence5(1,j)=P(q=qP(4 =i, q)(23-5)From the definitions of the forward and backward variables, E(,) and n(0), can be relateda, (Da, b(o)B() a,(ia,(o)B(j)(O1)∑∑a()ab(n,)Bn(y()=P(4=iO,)=∑P(q=,qn=jO,)=∑5(i,j)where >y,( denotes the expected number of transitions from phase i in O The terTherefore,λ(An be updated using(2-3-8)to(2-3-10)as followsAs Ti is the initial probability and denotes the expected frequency (number of times)inphase i at time t =1 as i=n1(), it can be calculated using the forward and backward()B(1)a1(1)A()(23-8)∑a1(i)B()∑a()∑5(,j)∑(23-9∑x()∑x()B()
A Scalable Healthcare Integrated Platform(SHIP) and Key Technologies for Daily Application 179Rapid changes in both societal environment and daily lifestyle are responsible for most ofall society and no longer just a private issue To elevate all citizensof thertance of healthd diseasto the steeplong-term healthere requirements, it is indispensable to be involved intake creative action A variety of innovative strategies and activities are now being explorednationwide in Japan at three levels: macro(administration), meso(community, organization,company), andlevel, a 12-year health promotion campaign, known as"Healthy Japan 21ealth promotion and fitness fetion,2000), has been advocated nationwidesince 2000 and is financially supported by the Japarministry of Health, Labour, ana"health promotion law" Japanese Ministry of Health, Labour andWelfare, 2002) was issued bgoal of medical insurance reconstruction was health promotion and disease prevention, andned individual responsibility and coordinat the meso level, industrial organizations and research institutions have developed mnet-based systems and related devices for daily healthcare Professional organizationof educationalprogramsaccreditation systems for professional healthcaretizen communities havePosted healAt the micro level, more and more people are aware of the importance of health promotionnd chronecomingparticipating in daily personalmedical examinations to keep their biochemical indices as good as possible chealthcare practices
They spendof time and meThis trend turns out that in the US only, the healthcare domain is now growing up intogiantindustrialterritoryworthyofaboutUs$2trillionannually(marketresearchcomof human welfare in long-term chronic treatment, weconfronting theding effective means for vital sign monitoring technologies suitable forlogical interconnection These solutions are being developeda terpretation of theirand large-scale data mining and a comprehensivethe world Manympanies are already engaged in and placing priority on, prThe"Health data bank asP servlatform was released as a multifaceted aid for thehealth management of corporate employee medical exam results(NTT Data Corp, 2002aidance and counselling, and takesngitudinal management in accumulated individual data Individual corporate employeesbrowse their personal data through Internet channels, andhs detailing historical changes in their health condition, thiilitaand research institutes in the European Union hdevelop wearable and portableHeart" project is a framework for personalPhilips, which aims to develop on-body sensors/electronics and appropriate services to help
Data Mining in Medical and Biological Researchght cardiovascular disease through prevention andsigns and physical movementand provide thewith recommendations (Philips Electronics, 2004) After thempletion of the"My Heart program, a continuation "HeartCycle project began in March2008 Many new sensors and key technologies, such as a cuff-less blood pressurewearable SpO2 sensor, an inductive impedance sensor, an electropuncture system,ontactless ECG, arrays of electret foils, a motion-comperperformance monitor(from bioimpedance) will be developed and built into the systempatients condition will be monitored using a combinationntient's clothing or bed sheets andd decision s ppliances, such as weighing scales and bloodpressure meters Data mining and decision support approaches willnort-term and long-term effects of lifestyle and medication,Mobihealtha mobile healthcare project funded by the European Commission from2002 to 2004 Fourt
eeners from hospitals and medical service providers, universinetwork operators, mobile application service providers, mobile infrastructure,ted in thetients to be fully mobile while undergoing health monitoring without much discomfort indaily activities The patients wore a lightweight unit with multiple sensors connected viaBody Area Network(BAN) for monitoring ECG, res? a need to stay in hospital(Europeanbody temperature using a wrist-mounted wearable device The device gathers the data a ngfheart rate, two-lead ECG, blood pressure, oxygen blood saturation, skin perspiration, andtransmits it to a remote telemedicine centre for further analysis and emergency care, using a(Anliker et al, 2004; European Commission, 2001)Healthvault"aibuild a universal hub of a network to connect personal health devicesd other services that can be used to helpande personal medical informationingle central site on the Web(Microsoft Corp 2008) It will provide a seamlessconnection interface scheme for various home health and wellness monitoring devices, suchas sport watches, blood glucose monitors, and blood pressure monitors marketed byufacturers worldOn the other hand, many expsIgnsmonitoring have been conducted in thecommercializedSince the first accurate recording of an ECG reported by willem Einthoven in 1895, and itsevelopment as a clinical tool, variants, such as Holter ECG, event ECG, and ECG mappingknown and have found a variety of applications in clinical practicchair(Lim et al, 2006)a toilet (Togawa et al, 1989), sleeps in a bed(Kawarada et al2000: Ishijima, 1993), sits in a bathtub(Mizukami et al, 1989; Tamura et al, 1997),en takesa shower(Fujii et al, 2002), his/ her heart beat can be monitored, with the person unawareThe smart dress, "Wealthy outfit, weaves electronics and fabrics together to detect thewearers vital signs, and transmits the data wirelessly to a computer The built-in sens
A Scalable Healthcare Integrated Platform(SHIP) and Key Technologies for Daily Application 18ement, eturethe suit looks and feelspletely normal (Rossi et al, 2008; Marculescu et al, 2003)The wellness mobilehas been山出小加m2081購加曲中a pedometer, a body fat meter, a pulse rate,According to an investigation report from the World Health Organization (WHO, 2002)ost current healthcare systems still have some comna) The difference between acute and chroniot sufficiently emphasized The ovoncept in system development has not shifted enough towards chronic conditions, and hasevolved to meet this changing demand(b) Despite the impobehaviour and adherence to improfor chronic conditioandsentialinformation to handle their condition to the best extent possible
(c) Patients are oftenlowed up sporadicalprovided with a long-term management plan forer thesesues a long-term difficult challengmmunities, and individuals aliketwoaspects should be paidttention The first aspegn monitoring for chronic conditions requiring different philosophy and strategy tends tobe ignored Long-term chronicfar from being " plug and plaG cious involvement in daily operation The second aspect is the lack of interconnectionetween multifarious physiological data within existingal systemsually decided by interpretation based on fragmented data and standards based on acutend emergent symptoms, and is often provided without the benefit of complete longmeet current needs, and to tackle the two problems abovestudies fpplicable domains, wherever vital signs are conducive2 Methods aopingof instrumental technologies and data miningthematical algorithms to construct finally a versatile platform, SHIIired andk technologies The following paragraphs describeerallvision of SHIP and introduce three related constitutional technologies thateveloping since 2002( Chen2004interpretation of long-term
Data Mining in Medical and Biological Researchysiological data using data mining mathematics), and (c)service(providing customizablety fields by a combinatAs shown in Fig 2, SHIP was constructed in a three-layer model which wapillaScalable Healthcare Integrated Platform(SHIP)[[lFig 2 Systemic architecture of the scalable healthcare integrated platform is founded onbricks, and supported by five pillars in a three layers structure A variety of applicationbe created using the SHIPThe first layer consists ofof bricks for physiological data detection in three ordersEach brick in the first ordeethod, wlindoors or outdoors, either awake or asleep
While eachbrick in the second and third orders indicates a data mining approach to defrom the other bricks The direct measurement signals are denoted as first-order vital signs,nd-order vital signs, such as heart rate, are drom the first-order parametersThird-oginate from the first- andbjects are: pressure, voice, gas, temperacceleration, plethysmogram, and urine The second-order vital signs are derived from thefirst-order vital signs, such as the Qrs widthrt rate from ecg, the pulse ratebreathing rate from the pressure( Chen et al, 2005), posture and body movement facceleration (Zhang et altransit time from the ECG andlethysmogram, The third-order vital signs are derived from both the second-order and thefirst-order vital signs For example, the variability in heart rate is derived from the heart rateprofile The female menstrual cyclmated from the bod2008a), and the variation in blood pressure is estimated from the prave transit time(Chen et al, 2000) Changes in these parameters are indicators of specific ailments,rrhythmia from the variability in heart rate,treat related illnesses such as cardiovascular disease, obesity andrespiratory obstruction
A Scalable Healthcare Integrated Platform(SHIP) and Key Technologies for Daily Application 18Data mining mathematics fotaond layeMost of stexploitation of the large volume ofm accumulated physiological data Innovativeerstanding frlaver will beThe third layer consists of three pillars and is responsible for data communication andbetween sensor devices and home units or mobile phones Mobile telephony and theInternettwo other pillars and used for wide -range data telation andSHIP has four characteristic features (a) Its ubiquity makes it possible to detect and collecttal signs either asleep (udoors or indoors through wearable/invisiblets and wired or wirelessks(b) Its scalabilityers to customizeecialmeetindividual needs, and also service providers to match different requests fromal/clinical use, industries, government agencies, and academic organizations throughpors) and guarantees that any emergentptured and responded to in real time
(d)15figured in two formats An exclusive format maintaindata transmission within ShIP, andthat SHIP is open to other allied systems through the HL7 standard(Health Levelseamless interface that is compatible with other existingmedical information systemsSHIP is intended to create a flexible platformthe exchange, management, andintegration of long-term data collected from a wide spectrum of users, and to provideSubjects in target servelderly andhealthcare but also subjects such as pharmaceutical houses fortherapeutic effect tracing,ansactions, transportation system drivers, fire fighters, and poecurityhitecture makesThree different tyfundamental instrumentation (invisible/ wearable/ubiquitous) and the results of datamining from our studies are introduced in the following sections22 Invisiblperforwav, such that a user is unaware of its existence and does not have to take care that theat aFig 3 Therelate and a bedside unit in the system configuration A sensor unit is placed beneathlow, which is stuffed with numerous fragments of soft comfortable materials formedfrom synthetic resins Two incompressible polyvinyl tubes, 30 cm in length and 4 mm inameter, are filled with air -free water preloaded tof 3 kPa and set in
Data Mining in Medical and Biological Researchparallel at a distance of 11 cm from each other A micro tactile switch(B3SN, Omron CoLtd) is fixed along the central line between the two parallel tubes The two tubes above andthe micro switch arached between two acrylic boards, both 3 mm thick One endch tube is hermetically sealed and the other end is connected to a liquiehead(AP-125, Keyence Co Ltd) The inner pressure in each tube includes static andynamIcts, andges in accordance with respiratory motion and cardiacbeating The static pressure component responds to the weight of the users head, and actsfttion of the ushead dbreathing movements and pulsatile blood flow fresure signals beneath thefar-neck occiput regions are amplified and bandpass filtered (0 16-5 Hz), and the staticponent is removed from the signal Only the dynamiconnection
The tactile switch is pressed to turn on a dC power supply via a delay switch(4387A-2BE, Artisan Controls Corp )when the user lies down to sleep and places his/herhead on the pillowside Uniternetlate is placed beneath a pillow Signals reflecting pressure changes under the piltected,digitized, and transmitted to a databasethe internet bdsideFig 4(a) The breathingR), heart rate(HR), and bodynentsdetected from the raw data measurementsThe BR and HR are detected by wavelet transformation on a dyadic grid plane using aprocedure, which is implemented by a recursive a trous algorithm The(CDF)(9,rthogonal wavelet is the basis function used togn the decomposition and reconstruction filters(Daubechies, 1992) The raw measurednks(Mallat Zhong, 1992; Shensa, 1992) Further mathematical theoriesan be found in Daubechies, 1992 and Akay, 1998 Implementation details are given in Chenet al 2005, and Zhu et al 2006
A Scalable Healthcare Integrated Platform( SHIP)and Key Technologies for Daily ApplicationThe wavelet transformation(WT) of a signal, x(t), is defined as followwhere s is the scale factor and y(r) is the wavelet basis function This is called a dyadic(EZ and z is the integral set) Two filter bankspass and high-passosition filters Ho and H, and associated reconstruction filters, go and Gt basis function and its scaling function, respectively Using Mallat'salgorithm, the dyadic WT of the digital signal, x(n), can be calculated as follorA1,x(n)=∑2=:A1x(k)(222)whereximation and detail components, respectivelythe 2' scale, and x(n)(or A, x(n))is the raw data signal The terms ho and h are the filtercoefficients of Hy and Hu, respectively Therefore, A, x(n)and D, x(beextracted from x()(or A, x(n) using equations (2-2-2)and(2-2-3) recursivescale approximation signal can also be reconstructed from the 2 scale approximation andx(k)+∑81n2D,x(k)here go andthe filter coefficients of Go and Gl, respectively
The terms x(n)A, x(n)can be finally reconstructed by repeatedly using equation(2-2-4) Any noised using a soft or hard threshold method beforereconstructed It should be pointed out that the sampling rate of the 2/ scale approximationnd detail is (/2, where fs is the sampling rate of theBecause the 26 scalreathing rhythm, whiltthe detail waveforms of both the 24 and 25 scales contain peaks similar to those of humeartbeats, the 26 scale approximation component, A6, is used to reconstruct the waveforcombined into a single synthesized waveform and then reconstructed to detect the HRFigure 4(b) shows the reconstructed waveforms for HR detection, and Fig 4(c) shows thereconstructed waveffor br detectieDuring a nights sleep, over a period of 4-8 h, a regular pulsation due to either the heartpressure variation signal pattern In such a time slot, either the Br or the HR, andmetimes even both, are barely detectable Instead, bodementsstatistical method in such time slots If a very large change, whose absolute value is fourtimes larger than the standard deviation of the preceding detected movement-free raw
Data Mining in Medical and Biolgnal, is detected in the incoming signal, the preceding and succeeding 25,s per nd frsdestimate the bR and HR Detection of the br is more sensitive to body metecti二解Fibodyt-free sample and the detected Br and hr beat-by-beat (a)The raweform reconstructed from the d4 and d5 components (c) the breath-related wayeftreconstructed from the A6 component The open circles indicate the detected characteristipoints for BR/HR determinationFigure 5(a) shows a 60 s segmentt duringnstable sleep When the pressure signal is distorted by a body motetected that are from two reconstructed waveforms(HR-related and BR-related)are notalways identical in both time and length
Because hr detection is usually more robust thanfrom hr-related waveforms The final bodyof both results In the case shown in Fig 5, the bodwaveform detection is counted as 154 s, while thatof br is 359 s thefinal bodent outputs as 37 8 s from the ORof both results in the timeFigure 6 shows a profile of the BR and the HR obtained fromments over a singlenight, The vertical axis denotes the BR/HR in units of breaths per minute or beats perminute(bpm) The black dots and vertical bars, terminated at the upper and lower ends byhes, show the mean values and standard deviaoccurring spond their widths denote periods of body movement The broader vertical bars correspond to