Automatic detection of respiratory rate from electrocardiogram, respiration induced plethysmography and 3D acceleration signals

本资料包含pdf文件1个，下载需要1积分

J．Cent．South Univ．(20 1 3)20：2423—243 1DoI：10．1007／sll771．0l3．1752一zAutomatic detection of respiratory rate from electrocardiogram，respiration induced plethysmography and 3D acceleration signalsLIU Guan—zheng(~ 'iE) ，WU Dan(吴丹) ，MEI Zhan．yong(梅占勇) ，ZHU Qing．song(~ ) ，WANG Lei(王磊) ，1．Biomedical Engineering Program，Stm Yat—sen University，Guangzhou 5 1 0006，China；2．Shenzhen Institutes ofAdvanced Technology,Shenzhen 5 1 8055，China；3．The Shenzhen Key Laboratory for Low-cost Healthcare，Shenzhen 518055，China；◎Central South University Press and Springer-Verlag Berlin Heidelberg 20 1 3垒SpringerAbstract：Respiratory monitoring is increasingly used in clinica1 and healthcare practices to diagnose chronic cardio—pulmonaryfunctional diseases during various routine activities．Wearable medical devices have realized the possibilities of ubiquitousrespiratory monitoring．however,relatively little attention is paid to accuracy and reliability．In previous study,a wearable respirationbiofeedback system was designed．In this work，three kinds of sign als were mixed to extract respiratory rate，1．e．，respirationinductive plethysmography fRIP)，3D．acceleration and ECG．In—situ experiments with twelve subjects indicate that the methodsign ificantly improves the accuracy and reliability over a dynamic range of respiration rate．It is possible to derive respiration ratefrom three signals within mean absolute percentage error 4．37％ of a reference gold standard．Similarly studies derive respiratory ratefrom single—lead ECG within mean absolute percentage error 1 7％ of a reference gold standard.

Key words：respiration inductive plethysmography；respiratory rate；electrocardiogram；3D acceleration；activity1 IntroductionNoninvasive respiratory monitoring has witnessed arapid surge of interest in recent years．The necessity forearly detection and diagnosis of chronic diseases，such assleep apnea [1]and COPD [2]，has fostered thedevelopment of difierent methods for measuringrespiratory activity,especially in ambulatory setings．Ingeneral， the capnograph devices such asB10PAC．MP 1 50 system with CO2 1 00C module[3]areroutinely used in hospital and healthcare centers，as thegolden standard for respiratory monitoring． Otherwell—accepted respiratory monitoring techn iques anddevices are electrical impedance tomography (EIT)，thermistors for airflow measurements and piezoelectrictransducers．Recently, non obtrusive respiration ratemonitoring methods such as respiration inductiveplethysmography (raP)， 3D—acceleration derivedrespiration rate fADR1 and ECG—derived respiration rate(EDR)，were increasingly demonstrated[4].

The appearance of the respiratory cycle in the heartrate signal(respiratory sinus arhythmia1 has been knownfor many years[5—1 0]．Many techniques have beendeveloped to derive respiration rate from electro—cardiogram (ECG)signa1．KHALED and FARGES[5]used an eighth order band—pass filter to derive respirationrate from ECG signa1 as early as 1 992．MASON andTARASSENK0 [6】used beat morphological feature toderive respiration rate．The wavelet transform has alsobeen widely applied to derive respiration rate from ECGsignal in clinic[7—9】．Furthermore，to find a measure ofrespiratory rate from a single—lead ECG recording thatwas the robust against noise from activities of dailyliving，B0YLE et al[1 0]compared performance amongsix diferent algorithms from band-pass filter and wavelet仃ansform．In this work．ECG derived respiration ratefEDR1 method was compared with respiration inductiveplethysmography (RIP) and acceleration derivedrespiration rate fADR)methods.

Foundation item：Project(2012M510207)supposed by the China Postdoctoral Science Foundation；Projects(60932001，61072031)supposed by theNational Natural Science Foundation of China；Project(2O12AA02A604)supposed by the National High Technology Research andDevelopment Program of China；Project(2013ZX03005013)supposed by the Next Generation Communication Technology MajorProject of National Science and Technology,China；Project supposed by the“One-hundred Talent’and the“Low—cost Healthcare’Programs of Chinese Academy of SciencesReceived date：2012——05——23；Accepted date：2012——09——18Corresponding author：WANG Lei，Professor，PhD；Tel：+86—15818518450；E—mail：wang．1ei###siat．ac．cn；liugzh3###mail．sysu．edu．cn2424 J．Cent．South Univ．(2013)20：2423—2431RIP is a 1atest respiration measure method．The RIPmethod consists of resistive bands that change electricalproperties based on the chest／abdomen wal movementsduring breathing【111．The RIP-based methods have beenproved to be an efective respiration monitoring method[12—14]．ZHANG et al[13]presented a RTP—basedwearable respiration monitoring device， which wasapplied in respiration biofeedback training．In previousstudy『14]，we have described a wearable wireless devicef0r monitoring different respiration activities such asrhythm，breathing mode and depth during sleep，based onthe digital RIP method．In this work．we aimed toestimate the robust performance of the R1P methodagainst noise from activities of daily 1iving.

The acceleration sensor has been also used to deftverespiration rate in recent years．Accelerometers worn onthe torso can measure inclination／angular changes duringbreathing，and then obtain respiratory rate[15—17].

HUNG et al[15]proposed a new approach based on achest biaxial accelerometer to derive respiratory rateduring static activities．BATES et al[1 6]designed a3D．accelerator from wireless sensor devices to deriverespiration rate．which仃acked the axis of rotation andobtained regular rates of breathing motion．ANMIN et al[1 7]suggested that the angle and principal componentsanalysis (PCA) methods could derive preciselyrespiratory rate from 3一D acceleration signals duringstatic activities．The aforementioned methods mainlyused spatial acceleration information．Therefore．whenthe subject was immobilizing，the respiration rateestimation inevitably deteriorated because the magnitudeof the movement．induced signal greatly exceeds that dueto breathing．and the posture and orientation of the 3．Daccelerometer were shifting during the disturbance．Inorder to ignore these problems，several authors simplyremoved the acceleration signal episodes that werecontaminated by motion artifacts『1 6—1 7]．However,inthis work． a novel spectrum analysis method waspresented to derive respiration rate from 3D accelerationsignals.

Versatile respiration rate detection methods havearoused widely atentions in recently years．To improvethe respiratory rate accuracy，DASH et a1[1 8]comparedthe performance of breathing rate detection algorithmsfrom three different physiological signals：the ECG，thephotoplethysmogram (PPG)and the piezoelectric pulsetransducer(PZO)signal during static activities．B0YLEet al f 1 01 took the lead in respiratory rate obtained fromone．1ead ECG during various routine (static anddynamic、 activities within mean absolute percentageerror 1 7％ of a reference gold standard.

A wearable device was introduced in this work tosense the ECG，the RjP and the acceleration signals【4].

In order to improve the accuracy and reliability,threesign als were mixed to derive respiratory rate across allactivities．Moreover,the diferences in performanceswere analyzed among the EDR，RIP and ADR methods.

2 System and methodsThe complete system was comprised of hardwaremodules for signal acquisition and processing algorithmsf0r respiration rate estimation．A waist．worn device wasdesigned to monitor RIP,ECG and 3D accelerationsignals based on the previous BSN platform developedbyLIU etalf41.

2．1 RIP method and EDR methodRIP method measures changes in rib cage andabdominal cross．sectional areas which are transiated intolung volume『19]．It has been an important method forrespiration signal monitoring．The EDR method usesvarious algorithms to obtain respiration rate from beatmorphology,heart rate，or a combination ofboth[14]．Inthis work．the RIP and ECG signals were monitored bywaist．wom device based on our general BSN platform .

The RIP method estimated of respiration rate from RIPsignal by using power spectrum method.

In our previous study,respiratory rate was derivedfrom ECG signals during static activities by using twoalgorithms based on beat morphology and RR interval，respectively『20]．There was no significant diferencebetween this method and other conventiona1 yetcumbersome methods to quantify the respiration rate，considering it is more accurate for wearable devices toobtain RR—interval than R amplitude． The powerspectrum analysis of the beat to beat heart rate variabilitywas chosen to derive respiration rate during staticactivities(e．g，siting，recover and specifc breathing).

However,the RR．interval cannot be detected duringdynamic activities，e．g，walking and running．Thus，a0．2-0．8 Hz Butterworth band．pass filter combining withpower spectrum [14]was chosen to derive respirationrate from the ECG signal during dyn amic activities.

2．2 Aeceleration derived respiration rate(ADR)The 3D acceleration signals were monitored usingthe waist-worn device during diferent routine activities.

In general，abdomen motion includes body motion andrespiration motion，and the frequency band for respirationvaries from 0．1 Hz to 0．6 Hz during various bodyactivities．Figure 1 demons仃ates the procedures thatrespiration rate is derived from abdomen 3D accelerationsignals during different routine activities． Firstly,according to one-minute energy expenditure(EE)，theband．pass filter was designed to obtain three respirationvectors from x．Y z coordinates of 3DJ．Cent．South Univ．(2013)20：2423—2431Adaptive digital filterTime／sTime／sTime／s(a)0 20Time／sTime／sTime／s2425Time／s Time／sFig·1 Step—by·step procedures for motion—derived respiratory rate estimation(a)，step—by—step signal processing results for real-worldsignals：raw 3D acceleration signals(b)，three vectors derived from 3D acceleration signals by using the band—pass filter based onone—minute energy expenditure(c)，respiration wave derived from three vectors by using the PCA method(d)and referencereparation wave from BIOPAC—MP 1 50 C02 module(e)One—minute energyZ expenditure( ) r— U 邑 >、兰 ∞0 0 矗薹g △0 C ．皇p∽
g 旨．莹 0 璺 8 宣盘 勺 ≥ 0 皇 WaVe．
旦p 0 一 > 0．J0 安．=0u 萝
．
喜差 ．J Q 1 『 1 r 1n §趸A buterworth band ∞ 8J，pass filter was chosen 一 △Z based on EE，u0_ 一Quu《0 O 0 。一 j9—0。。 ／u0一=l叠 一Du0《，u0 暑 0一 u0《002426 J．Cent．South Univ．(2013)20：2423—2431acceleration signals，respectively．Then，the principalcomponents analysis(PCA)was used to obtain theweights of every vector,and to extract the respiratorywave accurately．At last，respiration rate was computedby power spectrum analysis.

2．2．1 Energy expenditure algorithmAcceleration signals from three channels wereacquired at a sampling rate of 30 samples per second(Sps)．Acceleration of a moving object consisted of twopart：its own gravity and the acceleration caused byhuman movements， which were caled staticaccelerations and dynamic accelerations，respectively．Itis the latter that was concemed in our experiments.

Therefore，a high pass filter(一3 dB bandwidth 1 Hz)wasemployed to eliminate the static portions．Then，theacceleration signal A was defined as[2 1】4(f)=[( +1)一 )2+( +1)一 『)2+(z(f+1)一z(f)2]“ (1)The energy expenditure for abdomen part over oneminute was calculated asFrequency／Hz0．2 0．4 0．6 0．8 1．0 1．2= ∑ (f) (2)The routine activities were classified into threetypes based on energy expenditure as following：If EE<1 00，the activity was considered to be a lowEE；If 1 00SEE<400，the activity was considered to be amiddle EE；If EEl400，the activity was considered to be a highEE；2．2．2 Adaptive digital filterThe frequency of respiratory rate is approximately0．1 Hz to 0．6 Hz during different routine activities.

According to the diferent EE values of abdomen motion，the parameters of Buterworth band—pass filter wereadaptively selected to derive respiration vectors from ，Yand co—ordinates of 3D acceleration signals separately.

The parameters of BuRerwo~h digital filter are shown inFig． 2，whist the pass—band ripple Rp denotes themaximum permissible pass—band lOSS in decibels．and theFrequency／HzFrequency／Hz Frequency／HzFig．2 Power spectral density curves of four digital filters：(a)Parameters of first filter；Co)Param eters of second filter；(c)Param etersof third filter；(d)Parameters of fourth filter一 }{． 一／Jo 0【1 墨． ≯一／j 0一 IZ雹． 一／j0事0(1 N霉． 一，Jo事0J．Cent．South Univ．(20 1 3)20：2423—243 1 2427stop-band atenuation Rs denotes the number of decibelsthat the stop—band is down from the pass—band.

To derive the respiration vectors，the adaptivedigital filter was chosen based on the one-minute energyexpenditure(EE)as folowing：1)When the energy expenditure is high，theband—pass filter was automatically chosen torespiration vectors from 3一D acceleration，such asrunning at 6km／h.

secondderiveduring2)When the energy expenditure is middle，the thirdband．pass filter was automatically chosen to deriverespiration vectors from 3．D acceleration such as duringwalking at 2 km／h.

31 When the energy expenditure is low,the fourthband—pass filter was automatically chosen to deriverespiration vectors from 3-D acceleration，such as duringspecific breathing，siting and recover.

The parameters of Butterworth band—pass filterdenote as"X=[frequency band，stop band，Rp，Rs] (3)2I2．3 Principal component analysisBecause of the unpredicted posture changes duringroutine activities and the geometric deployment of theaccelerometer，a PCA-based method was proposed toobtain the weight of the respiration vectors from the ，Y，an d x axes accelerations.

Principal component analysis(PCA)is generalyused for dimension reduction of multivariate datasets[22].

Firstly，denoting the three co-ordinate accelerationvectors time series as matrix ={ (k)； (k)； (k)}.

Secondly， the eigenvectors and correspondingeigenvalues( ， ， )of was computed based onthe PCA method[221.

Thirdly， the weights for three co—ordinateacceleration vectors were obtained as，7f=by1+ 2+23 “=1,2，3) (4)At last，x0 represents the respiratory time sequencex0 T]1Vx+q2Vy+r]3vz (5)where the vx，Vy and vz denotes e vectors from x，Y and Zco—ordinates of 3D acceleration filtered using adaptiveband．pass filter.

2-2．4 Respiratory rateRespiratory rate was computed by using powerspectrum．The fast Fourier transform is ca~ied out usinga sliding window for l min；and the window length forpower spectrum is 1 800 sampling points．The slidingwindow has a length of60 s.

2．3 M ixed respiratory rate derived methodAccording to the features of three respiratory ratederived methods a mixed three signals method was takenas follows：1) Wh en the activity is specific breathing，respiratory rate was obtained by RIP method.

2)When the activity is siting and recover(1ow )，respiratory rate was obtained by combination of EDRmethod and RIP method.

3)Vv~l_en the activity is waking (middlerespiratory rate was obtained by combination ofmethod and RIP method.

4) When the activity is running (highrespiratory rate was obtained by ADR method.

3 Experiments)，ADR1E)，3．1 Protocol of exoerimentIn total，twelye healthy subjects with a mean age of25．5(between 21 and 32)participated in the experiments，and the protocol was given in Tlable 1．Al1 subjects wereconsented on a voluntary basis an d the study wassubjected to internal review．伟 en subjects were askedto breath at the fix frequency(specifc breathing)，amusic metronome was used to regulate his／her breaths atfixed respiratory rates(i．e．，8，12，16，20，24 breaths perminute).

Table 1 Experimental protocolThe con仃olled 30 min recordings followed theprescribed protocols as shown in Fig．3：The reference of respiration(golden standard1 wasobtained by using MP1 50 from BIOPAC systems Inc—with C02 1 00C analysis module．Then．the respirationrate was computed by using power spectrum method.

2428 J．Cent．South Univ．(20 1 3)20：2423—243 1Sign
．
al andetimeReference signalECG signal ’ ? ?? ? ? ——'7r— — ——一’。 ’一’一呵— — — — — — ． ， ． — — “ — — — 一 RIP sign al蠡_山 誓 黼 秤即 F- 霸 0黼—r? ? ’ r? ? A
cceleration． 一 一 ～ ． ． 。
_删糟嘲I曩i瞄 嘲嘲—哪煽 ?。 ．i,-isignalsTime／min 0 5 10 15 20 25 30Fig．3 Raw data for 30 min controlled respiration studies(Golden standard means reference respiratory signal by using MP 1 50 fromBIOPAC systems Inc．，with C02100C analysis module；ECG,RIP,3-D acceleration ，Y，z)signals was obtained by using ourwaist．worn device3．2 Statistical analysisThe three methods(RIP．EDR．and ADR methods)were employed and the accuracy of each method wasexpressed by determining the average percentage error ofthe measurements when compared with the referencevalue(golden standard)，calculated by subtracting thetest value from the reference value and dividing theresults by the reference value of each measurement，andthen took the average of this『41．Absolute value wasused to avoid eror offsetting between each other.

： I 二 100 (6)where E is the error：VD is the derived value； is thereference value；Vo is the observed value.

The statistical tests (with SPSS vl 7．01 wereconducted for al the tests．Mean and standard deviationswere used to evaluate the mean absolute error betweenthe derived results and the reference values．Whel1 tw odatasets were compared， tw o sample t-tests wereperformed for each individual method．The signifcancelevel chosen was a=0．05.

4 ResultsA complete comparison for the three methodsconsidered required statistical comparisons of detectionaccuracy and consistency for a variety of breathingconditions．At last，according to those features，a highaccuracy respiratory rate was obtained with the mixedrespiratory rate method.

4．1 Overall oerformance of all aetivitiesFigure 4 demons仃ates the Bland—Altman of threemethods across all activities．The results indicate thatRI P method with erors smaller than 5 breathes perminute can be used to respiration rate monitoring duringall activates．Figure 5 depicts the mean absolute erors ofthree methods betw een the derived results and thereference respiration signals for all activities during30 rain controlled test．The RI P method has the leastmean absolute eror with the reference signal at 4．63％across all activities tested．Multiple comparison testingindic：ates that there is no significant diference in thereference respiration rate betw een the EDR and ADRmethods．Thus．considering overall perform ance acrossal activities，the RI P method，which has the bestrobustness，is the best respiratory rate monitoring method；and the EDR and ADR methods give equivalentperform ance in deriving respiration rate.

4．2 Differences among routine activitiesThe dataset was partitioned into the followingJ．Cent．South Univ．(20 1 3)20：2423—243 1 2429420矗勺 ‘一 2- 420151050- 5一 l0— 152020(a)+1．96 SD一 。 2
．5oo一 o
。
o ∞三 ∞ 。 。。 M ean。 o o o 。

一 0．4。 。
o 。o 。

一 o u o o。 。 一1．96 SDo 一3_3o o1O 15 20 25 30 35 40 45Average of RIP and golden standard(b) 。

o +1．96 SD。 。
。o。

。 。

。 11·9一 。

一
。

。
。o
。 M ean。 如 一。 。 。 。 1
．9o o— o oo 一1．96 SD一
? ? ? ? ? ? ? ? ? ? ? ? ? o] l l _ ● l l ●5 l0 15 20 25 30 35 40 45Average of EDR and golden standard(。) 十1
． 96 SD。 l0．7o o一 _。c 。。 。 8。。。 M。

。 。 ．3一 。

。 一1．96 SD?????‘。 ???????? 芝‘10 15 20 25 3O 35 40 45Average of ADR and golden standardFig．4 Bland—Altman plots of three methods across al1 activities(specific breathing，siting，walking，running and recover)：Ca)Bland—Altman plot between RIP and golden standard；(b)Bland—Altman plot between EDR and golden standard；(c)Bland—Altman plot between ADR and golden standard(RIPmeans digital respiratory inductive plethysmography derivedrespiration rate；EDR means ECG—derived respiration rate；ADR means tri—acceleration derived respiration rate)2520呈∞ 100RIP EDR ADRFig．5 Mean absolute errors of three methods during alactivities(specifc breathing，siting，walking，running andrecover).

activity categories：sitting，walking，running，recover andspecific breathing．Figure 6 indicates the differencesamong three methods for each activity tested．Firstly，forthe low EE activities(e．g．，sitting，recover and specificbreathing)，the ADR method has the largest meanabsolute error with the reference signal at above 1 0％：however，the ADR method has the least mean absoluteerror during dynamic activities such as walking andrunning．Secondly，the EDR method has the largest meanabsolute error with the reference signal at the beyond25％ during dynamic activities such as walking andrunning．but it has the least mean absolute eror with thereference signal at below 5％ during siting．Third．theRIP method has the 1east mean absolute error with thereference signal at below 5％ during recover and specificbreathing.

Therefore，the most obvious diference is that theEDR method is easily affected by various noises.

especially for motion artifact during dynamic activities(e．g．，walking and running)；the ADR method is robustagainst motion artifact.

4．3 M ixed three signals respiratory rate derivedm ethodAccording to the diferences of three respiration ratederived methods，the mixed respiratory rate derivedmethod makes fu1 use of the advantages of each method.

For a range of common activities of daily lire tested inthe study(specifc breathing，siting，walking，runningand recover)．it is possible to derive respiration ratewithin mean absolute percentage error 4．37％ of areference gold standard．By the way,it improves therespiration rate monitoring reliability.

口j1 口 ∞ 蛊∞Iop10 Q0 0 0 0 0 口_I对 ∞=I∞IlQ 一0 ． n【'
J．Cent．South Univ．(20 1 3)20：2423—243 1 2431energy expenditure by our wearable medical deviceReferences『1] YOUNES M．Role of respiratory control mechanisms in thepathogenesis of obstructive sleep disorders[J】_J Appl Physiol，2008，l05：l389-1405.

[2】 PATEL S，MANC1NELLI C，BONATO P，HEALEY J，MOY MUsing wearable sensors to monitor ph3，sical activities of patients withCOPD：A comparison of clasifier performance[J]．Wearable andImplantable Body Sensor Networks，2009：234—239.

[3] Ⅵnvw．biopac．corn【4] LIU G Z，HUANG B Y，WANG L．A wearable respiratorybiofeedback System based on generalized body sensor network[J].

TelemedicineandE-health，2011，l7(5)：348—357.

[5] KHALED z，FARGES G．First approach for respiratory monitoringby amplitude demodulation of the electrocardiogram 【Cll Proc 14thAnnu Int Conf IEEE Engineering in Medicine and Biology SocietyParis，France，1992：2535—2536[6] MASON C，TARAssENKO L．Quantitative asessment ofrespiratoryderivation algorithms rC]／ Proc 14也 Annu Int Conf IEEEEngineering in Medicine and Biology Society．Istanbul，Turkey，2001：1998-2001.

[7] ROCHE F，PICHOT V，SFORZA E，COURT—FORTUNE I，DUVERNEY D．COSTES F．GARET M BARTHELEMY J C.

Predicting sleep apnoea syndrome from heart period： Atime-frequency wavelet analysis[J]．Eur Respir J，2003，22：937—942.

[8] RAYMOND B，CAYTOM R，BATES R，CHAPELL M．Screeningfbr obstructive sleep aDnoea based on the electrocardiogran a- Thecomputers in cardiology challenge[c]／27th Annual Meeting ofComputers in Cardiology．Cam bridge，MA，2000：267—270.

【9] SCHRADER M，ZYWIETZ C，von EINEM V，WIDIGER B，JOSEPH G．Detection of sleep apnea in single channel ECGs from thePhysioNet data base[c]／27th Annual Meeting of Computers inCardiology．Cambridge，M A，2000：263—266.

[10]BOYLE J，BIDARGADDI N，SARELA A，KARuNANITHI M.

Automatic Detection of Respiration Rate From AmbulatorySingle Lead ECG [J1．IEEE Trans Information and Technology inBiomedicine，2009，13：890—896.

[11】MACKDC，PATRIE JT，SURATTP FELDERRA，ALWANM.

Development and preliminary validation of heart rate and breathingrate detection using a passive，ballistocardiography—based sleepmonitoring system[J]．IEEE Transactions on Inform ation Technologyin Biomedicine，2009，13(1)：11 1—120.

[12]NAM S H，YIM T C，et a1．The preliminary study of unobtrusiverespiratory monitoring for e-health[C]／27th Anual InternationalConference ofthe IEEE Engineering in Medicine and Biology Society.

Shanghai，China，2005：3796—3798.

[13]ZHANGZB，WANGWD，WANGBQ，wuH，LIUHY，ZHANGY K．A prototype ofwearable respiration biofeedback platform and itspreliminary evaluation on cardiovascular variability [c]／ 3rdIntem ationa1 Conference on Bioinform atics and Biomedica1Engineering(ICBBE 2009)．Beijing，China，2009：1397—1400.

[14]wu D，WANG L，ZHANG Y T．A wearable respiration monitoringsystem based on digital respiratory inductive plethysmography[c]／c0nf 2009 IEEE Engineering in Medicine and Biology Society．2009：4844-4847.

[15]HUNG P D，BONNET S，GILLEMAUD R，CASTELLI E，YEN P TN．Estimation of respiratory waveform using an accelerometer[c]／30th Annual International Conference of the IEEE Engineering—-in·Medicine—and—Biology—Society． Vancouver， Canada， 2008： 1493—1496.

[16]BATES A，LING M J，MANN J，ARVIND D K．Respiratory rate andflow waveform estimation from tri-axial accelerometer data【c]／2010International Conference on Body Sensor Networks(BSN)．Singapore，Singapore，2010：144—150.

[17]ANM1N J，BIN Y，GEERT M，HARIS D Performance evaluation ofa tri-axial accelerometry-based respiration monitoring for am bientassisted living[c]／31 st Annual International Conference ofthe IEEEEngineering in Medicine and Biology Society． EMBC 2009，Minneapolis，MN，USA，2009：5677—5680.

[18]DASH S，SHELLEY K H，SIVERMAN D G，CHON K H．Estimationof respiratory rate from ECG，photoplethysmogram，and piezoelectricpulse transducer signals：A comparative study of time-frequencymethods[J]．IEEE Trans Biomedical Engineering，2010，57(5)：[19]FIAMMA M N，sAMARA z，BAcONNIER P，SIMILOWsK T，STRAUS C．Respiratory inductive plethysmography to assessrespiratory variability and complexity in humans[JJ．RespiratoryPhysiology &Neurobiology，2007，156(2)：234—239.

[20]wu D，LIU G Z，WONG M Y M，ZHANG Y T．The accuracy ofrespiratory rate estimation using electrocardiography andphotoplethysmography[c]／10th IEEE International Conference onInformation Technology and Applications in Biomedicine(ITAB2010)．Corfu．Greece．2010：1_4.

[21]LIN S J，WANGL，HUANGBY，ZHANGYT，wuxM，ZHAO JP．A pilot study on BSN—based ubiquitous energy expendituremonitoring[c]／Conference of 2009 IEEE Body Sensor Networks，2009：1_4[22]DUDA R O，HART P E，STORK D G．Patern classifcation[M]．2ndEdition．Beijing：China Machine Press and Citic Publishing House，2003：568-569.

(Edited by HE Yun—bin)