patwari11breathing.pdf

8/10/2019 patwari11breathing.pdf

1/10

arXiv:1109.3898v1

[cs.NI]18Sep2011

ARXIV.ORG TECHNICAL REPORT 1

Monitoring Breathing via Signal Strength inWireless Networks

Neal Patwari, Joey Wilson, Sai Ananthanarayanan P.R., Sneha K. Kasera, Dwayne Westenskow

AbstractThis paper shows experimentally that standard wireless

networks which measure received signal strength (RSS) can be used

to reliably detect human breathing and estimate the breathing rate,

an application we call BreathTaking. We show that although an in-

dividual link cannot reliably detect breathing, the collective spectral

content of a network of devices reliably indicates the presence and

rate of breathing. We present a maximum likelihood estimator (MLE)

of breathing rate, amplitude, and phase, which uses the RSS data

from many links simultaneously. We show experimental results which

demonstrate that reliable detection and frequency estimation is possible

with 30 seconds of data, within 0.3 breaths per minute (bpm) RMS error.

Use of directional antennas is shown to improve robustness to motion

near the network.

1 INTRODUCTION

In this paper, we explore the ability to detect and monitorbreathing using the changes in received signal strength(RSS) measured on many links in a deployed wirelessnetwork. The ability of a wireless network to makemeasurements that can monitor a persons breathing cancreate new opportunities for improving patient moni-

toring in health care applications. As one example, post-surgical patients can die from respiratory depression andairway obstruction, which are unfortunately commonafter surgery due the difficulty of correctly dosing seda-tives and pain medications administered to a patient [1].Reliable respiration monitoring is critical to detection ofthese conditions [2], [3], [4]. Breathing monitoring alsohas application in diagnosis and treatment for obstruc-tive sleep apnea, in which a person experiences periodsof low breathing rate or long pauses in breathing whilesleeping [5]. Finally, breathing monitoring may haveapplication in detecting sudden infant death syndrome(SIDS), which is one of the largest causes of death in

infants. Parents with a child at high risk for SIDS maywish to use a baby breathing monitor to alert them incase their childs breathing becomes depressed or stops.

We use of measurements of RSS between many pairsof wireless devices in a deployed network to non-invasively detect and monitor a persons breathing, an

N. Patwari is with the Department of Electrical and Computer Engineering,University of Utah, Salt Lake City, USA. J. Wilson is with Xandem Technol-ogy, Utah, USA. S.K. Kasera is with the School of Computing, University ofUtah. S. Ananthanarayanan is with Motorola Mobility, USA. D. Westenskowis with the Department of Anesthesiology, University of Utah. This materialis based upon work supported by the National Science Foundation underGrant Nos. #0748206 and #1035565. Contact email: [email protected].

application we call BreathTaking. While severe fading inmobile radio channels is expected, it is counterintuitivethat small changes in a persons size as a result oftheir breathing could be detected using measurementsof RSS. However, in this paper, we demonstrate thatin an otherwise stationary environment, when we usethe data collected on many links between static wireless

devices, breathing monitoring is not only possible, butremarkably reliable.

0.2 0.3 0.4 0.5 0.630

40

50

60

70

80

90

100

110

Frequency (Hz)

N

ormalizedAveragePSD

Norm. Avg. PSDActual Breathing Rate

Fig. 1. Normalized averaged power spectral density(PSD) vs. frequency (Hz). Total PSD is defined as theargument of (4). The peak of the PSD plot is at 0.253Hz (15.18 bpm), compared to the actual breathing rate of

0.250 Hz (15 bpm), shown as a vertical dashed line.

Our research on BreathTaking is motivated by exper-

imental observations. We have observed that when aperson is standing on or near the line-of-sight (LOS)of a static radio link, the RSS can be changed simply

by the persons inhaling and exhaling. In fact, when wemeasure the actual breathing rate and analyze the powerspectral density of the RSS link data from the networkas a whole, as shown in Figure 1, we see a peak veryclose to the actual breathing rate.

Our BreathTaking does not provide a direct measure ofbreathing. In contrast to End-Tidal CO2 monitoring, forexample, we do not measure the gasses exhaled from apersons nose. We simply make an observation about the
http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1http://arxiv.org/abs/1109.3898v1


2/10


presence (or absence) of a strong frequency componentin the measured RSS in the human breathing range.Adults, at rest, breathe at about 14 breaths per minute(bpm)[6], while newborns breathe at 37 bpm [7]. To beinclusive, we consider a range of 10 to 40 bpm (0.167to 0.667 Hz). Few other objects have cyclic motion withperiodicity in this range, but if there were such an objectin the deployment region, it might cause the same typeof observation in the network RSS data. Context remainsimportant to interpret results from the proposed system.

In this paper, we make the following important con-tributions in relation to system design, capabilities, andlimitations of BreathTaking. First, we develop methodsto accurately estimate the breathing rate and reliablydetect breathing of a person in the deployment areaof a wireless network by considering the RSS mea-surements on many links simultaneously. We approxi-mate the breathing signal to be sinusoidal and use themaximum likelihood estimation (MLE) to estimate the

breathing rate. Second, we perform extensive experimen-

tal evaluation of BreathTaking in an indoor setting. Wedemonstrate monitoring the breathing of an otherwisemotionless person, in an hospital room with no otherperson present. With thirty seconds of RSS data in atwenty-device network, we demonstrate (1) breathingrate estimates with RMSE between 0.1-0.4 breaths perminute (bpm); (2) a breathing detector without falsealarms or missed detections during experiments per-formed with devices connected to directional antennas.We address the performance as a function of the numberof devices in the network, relative position of the personwith respect to the sensors, and actual breathing rate.

Finally, we quantitatively address the following key

questions that relate to the capability to use RSS mea-surements from static wireless networks to monitor

breathing for the above-described applications:

1) What is the benefit of using data from multiplelinks simultaneously, as opposed to from one link?

2) How accurately can breathing rate be estimated?3) How long of a measurement duration is required?4) What is the effect of the directionality of the anten-

nas?5) How many devices are required for accurate mon-

itoring?6) Is there information in the phase of the breathing

signal?Breathing monitoring in a wireless network has ap-

plications and implications besides health care. A searchand rescue team may arrive at a collapsed building andthrow transceivers into the rubble, hoping to detect the

breathing of anyone still alive inside. Police or SWATteams may deploy a network around a building todetermine if people are inside. On the other hand, theability to measure breathing from a wireless networkhas privacy implications. We have shown previouslythat a network deployed around the external walls ofa building can detect and track a person who is moving

or changing position [8], [9]. If this system can also detecta persons breathing, it can also detect people who aresitting or laying motionless.

The rest of the paper is organized as follows. In Section2 we present the approach we have used for Breath-Taking. Next, we describe the experimental testbed andmethodology in Section3.The results of the experimentsare presented and discussed in Section4. Finally, relatedwork and conclusions are presented in Sections5 and 6.

2 METHODS

In this section, we define the measurements, models, andgoals of the BreathTaking system.

2.1 Network

We assume a network in which received power (alsocalled RSS) measurements can be made on many links

between pairs of wireless devices. We assume thesemeasurements can be made often, at regular intervals.

Specifically, assuming a maximum breathing rate of 40breaths per minute, BreathTaking requires RSS measure-ments to be made at a rate higher than 4/3 Hz, theNyquist sampling rate.

We denoteyl(i)to be the dBm received power on linkl measured at time i. Note that link l is an ordered pair(tl, rl) of the particular transmitter tl and receiver rl forlinkl. We do not generally require full connectivity of thenetwork, and instead, assume that connected links arenumbered from 1 throughL, whereLis the total numberof measured links. We wish to maximize L and thus usea wireless sensor network with a mesh topology in ourexperiments, although we do not exclude networks withother topologies.

2.2 Signal Model

In the absence of any motion in the environment of thenetwork, we denote

yl(i) = yl+l(i), (1)

where yl is the mean RSS for link l and l(i) is additivenoise. We assume that the noise on link l is i.i.d. zero-mean Gaussian. We also assume that the noise l(i) isindependent on different links l. In the presence of a

breathing person, we assume that the link RSS has anadditional sinusoidal term,

yl(i) = yl+Alcos(2f Tsi+l) +l(i), (2)

where Al, l, and f are the amplitude, phase, andfrequency, respectively, of the periodic component of theRSS signal on link l, and Ts is the sampling period. Weassume that the periodic component due to breathingwould have the same frequency on all links l, and thatthe sampling period is made to be identical on all links,thus we do not use a subscript l for frequency f. Phaseand amplitude are expected to differ between links.


3/10


2.3 Framework

We denote the measured signal on link l as a vector,

yl = [yl(1), . . . , yl(N)]

where N is the total number of samples, and () indi-cates the vector transpose. Because the sampling periodis Ts, the measurement vector corresponds to what we

call the observation period T,T=N Ts.

The observation period T is related to the latency ofbreathing monitoring, since we measure RSS for a dura-tion Tbefore obtaining estimates of breathing rate anddetecting whether or not breathing occurred.

Our objective from the measurements yl, for l =1, . . . L, is to detect whether or not a person is breathingwithin the network, and to estimate important parame-ters of the model, which we denote ,

= [AT, T, f]T

where A = [A1, . . . , AL]T and = [1, . . . , L]T. Ofmost interest is the frequency f, as human breathing hasa characteristic frequency range, from fmin to fmax, asdiscussed in the Introduction. In this paper, we detect

breathing only within fmin = 0.167 Hz to fmax = 0.667Hz, and in general, we assume that the range is given.

2.4 DC Removal Filtering

When estimating the power spectral density of noisy,finite-duration yl signals, the mean values yl (the DCcomponent) can hide the power of lower-amplitudesinusoidal components. Yet the DC component doesnot hold information about the presence or absence of

breathing. Since we assume that breathing is not presentbelow a frequency offmin, we address this problem sim-ply by using a high pass filter which strongly attenuatesthe DC component. We do not require a linear-phasefilter, but we do need a nearly flat amplitude gain abovefmin, since we do not want our system to bias towardssome frequencies because they have been amplified by afilter ripple. In all results in this paper, we use a 7th orderChebychev high-pass filter with maximum passbandripple of 0.1 dB and passband frequency fmin = 0.167Hz. An order of 7 was found to be sufficient to have atleast 40 dB attenuation at frequencies lower than0.1Hz.

For all further discussion, we assume that each linksyl(i) signal has been filtered using this high-pass filter,and thus yl = 0.

2.5 Breathing Estimation

Breathing frequency estimation plays a primary rolein breathing detection, and thus we discuss frequencyestimation prior to breathing detection. In this section,we present the maximum likelihood estimate (MLE) of

breathing parameters, including frequency, link ampli-tudes, and link phases, given the model presented inSection2.3.

Maximum likelihood estimation of is an extension ofthe standard sinusoid parameter estimation problem [10,p. 193195] in which there is a single signal composedof one sinusoid of unknown phase, amplitude, andfrequency in additive white Gaussian noise. In our case,we additionally have L different link signals, each withits own amplitude and phase, and we have a frequencylimited to the range [fmin, fmax].

In our case, under the i.i.d. Gaussian noise modelin the presence of breathing, the likelihood function ismaximized when the following functionJis minimized:

J() =Ll=1

Ni=1

[yl(i) Alcos(2f Tsi+l)]2 (3)

One can modify the derivation of [10, p. 193195] forthis case, that is, to minimize J() in (3), and show thata good approximation of the MLE of frequency fis given

by

f= argmaxfminffmax

Ll=1

Ni=1

yl(i)ej2fTsi2

(4)

The approximation is very good whenever the normal-ized frequency, f Ts, is not very close to 0 or 1/2. In ourcase, we specifically exclude frequencies close to zero,and sample at a frequency significantly higher than theNyquist rate.

Note that if one wishes to estimate breathing fre-quency from one link alone, one may use (4)withL = 1.

The maximum likelihood link amplitude estimates{Al} and phase estimates {l}are then estimated usingf, and are given by

Al = 2

N

Ni=1

yl(i)ej2fTsi

(5)

l = arctanN

i=1yl(i)sin2f TsiN

i=1yl(i)cos2f Tsi

.

2.6 Breathing Detection

We consider deciding between two hypotheses:

H0 : A breathing person is not present (6)

H1 : A breathing person is present (7)

For detection, we study two methods:

1) Single-link breathing detection: Use solely the RSSmeasured on one link in order to detect breathing.

2) Network-wide breathing detection: Use the RSS mea-sured on all L >1 links in the wireless network todetect breathing.

By comparing the two methods, we quantify the im-provement in breathing detection possible when datafrom many links in a network are used.


4/10


2.6.1 Single-link breathing detection

Consider one links RSS data. Without loss of generality,assume the link number is l = 1. Then the MLE offand Al are calculated from (4)and (5) using L = 1. Oursimple assumption is that Alwill have higher amplitudewhen a breathing person is present. Thus we detect

breathing via the hypothesis test,

NA2lH1> 0.72, 17 provide f estimateswithin 1 bpm of the actual breathing rate (15 bpm).However, the other 46 estimates are nearly uniformlydistributed on [fmin, fmax].

In sum, one cannot expect to detect breathing orestimate breathing rate on any single deployed link.Further, even if one has many links in a network, itwould be unreliable, as a monitoring method, to lookat single-link amplitude and frequency estimates.

4.2 BreathTaking Rate Estimation

Next, we consider at network-wide breathing monitor-ing. For the sameH1experiment described in Section4.1,we study network-wide breathing rate estimates, whichare calculated using (4) with L set to the total numberof links in the network. We calculate rate estimation

performance for a variety of different periods T. Thevast majority of breathing rate estimates fall within 5

bpm of the actual rate a small fraction do not. We callthese estimates that are more than 5 bpm from the actualrate invalid rate estimates, and report the percentageof rate estimates that are invalid. We also calculate theRMSE and bias of the estimates that are valid, i.e., within5 bpm of the true rate.

The experimental results, shown in Figure 3, showthat for T 30s, less than 2% of rate estimates can

be described as invalid. The RMSE for valid estimatesis lower than 0.5 for all observation periods T 25s.


6/10


20 30 40 50 600.2

0

0.2

0.4

0.6

0.8

1

1.2

1.4

BreathingRateError(bpm)

Observation Period T (s)

20 30 40 50 600

2

4

6

8

10

12

14

16

PercentofEstimatesInvalid

RMSE

Bias

Fig. 3. RMSE and bias of valid BreathTaking frequencyestimates for controlled breathing / patch antenna exper-iment (actual breathing rate of 15 bpm), and percent of

rate estimates that are invalid (right y-axis), vs. observa-tion periodT.

For perspective, current medical devices typically re-port breathing rate as an integer number of breathsper minute, thus rate errors less than 0.5 bpm would

be insignificant. The bias is small, on the order of 0.1bpm. For T 50s, there are no invalid breathing rateestimates.

4.3 BreathTaking Amplitude Estimation

Once we obtain the MLE of frequency f using thenetwork RSS data as in Section4.2, we can estimate the

amplitudes Al. Note that there are no actual values ofAl; some links will measure high amplitude, and others

will not. We are particularly interested in the links lwith particularly high Al. For the same H1 experimentdescribed in Section 4.1, consider the Al for T = 30 s.Only 5% of links l have an amplitude Al > 0.331, linkswe refer to as high amplitude links.

In Figure4,we plot the locations of the high amplitudelinks by drawing a dashed line between their transmitterand receiver coordinates. We can see that the linkswhich cross through the chest area are the ones that areparticularly affected by breathing.

Still, only a fraction of the links that cross through

the chest measure high Al. What other characteristicsdo these high amplitude links have? We find that highamplitude links have unusually low average RSS. Overall links, the average measured RSS is -39.9. Whenconsidering only high amplitude links, the average mea-sured RSS is -48.6, almost 9 dB lower, a very significantdifference. This difference cannot be explained by longerpath length high amplitude links are only 13% longer,on average, than the average path length of all links.Since links that in a deep fade experience greater tem-poral variations due to changes in the environment [9], itmakes sense that the amplitude of the breathing-induced

0 0.5 10.5

0

0.5

1

1.5

2

0 1 2 3

4

5

6

7

8

9

10111213

14

15

16

17

18

19

X Coordinate (m)

YCoordinate(m)

Bed Outline

Person

6 SensorCoord & No.

High Ampl.Link

Fig. 4. Experimental layout showing sensors, bed, and

persons approximate position. Dashed lines indicatehighamplitude links.

change would be more noticeable for links with lowerthan average RSS.

One lesson is that links in a deep fade are more use-ful for breathing monitoring. Future work may explorechanging the center frequency on links, or measuringwideband frequency response, with the goal of adap-tively using data from links frequency nulls. Althoughwe do not explore this idea in this paper, we wouldexpect to be able to improve results by taking advantageof deep fades wherever in the frequency spectrum theyoccur.

4.4 BreathTaking Detection Performance

In the network-wide case, breathing detection is per-formed using a normalized sum of the squared am-plitudes |Al|2 over all links l, as given in (9). In thissection, we study the performance of this detector. Usingthe same experiments as presented in Section 4.1, wecalculate S from (9) for each T second period, testingdetector performance for each T in the range 15 to 60seconds, in 5 second increments. Figure 5 shows theprobability density functions (pdfs) ofS for H0 and H1cases, for T = 15 (top subplot) and T = 30 (bottom

subplot).We find first that S in the H0 case always falls in a

narrow range, between 0.98 and 1.45. During H1, theS values has a minimum of 1.57 (at T = 15). Further,as T increases, Svalues also increase. For T = 30 andT = 60, the minimum S values recorded are 1.63 and2.03, respectively.

We can conclude that for this experiment, because ofthe lack of overlap in value ofS in the two cases, onecan build a reliable detector. For example, we can setnet = 1.50 in (9) and perfectly distinguish between noperson present and a breathing person present.


7/10


1 1.5 2 2.5 3 3.5 4 4.5 50

2.5

5

7.5

Prob.

Density

T= 15s Network-wide Breathing Statistic S

1 1.5 2 2.5 3 3.5 4 4.5 50

2.5

5

7.5

Prob.

Density


Empty

Person

Empty

Person

Fig. 5. Probability density functions for S given H0(Empty) vs. given H1 (Person) for the (top) T = 15and (bottom) T = 30 second observation periods, for

experiments using patch antennas.

4.5 Antennas

Experiments discussed above use patch antennas witha 8 dBi antenna gain. In this section, we compare theresults when using less directional antennas. We runmore experiments with all nodes connected to dipoleantennas, which have a omnidirectional pattern in thehorizontal plane. The first experiment is an H0 experi-ment, with no person in the room, and little movement inthe hallway outside of the room. The second experimentis an H1 experiment, with the person lying in the bedand breathing at a constant rate of 15 bpm (using ametronome).

The results, shown in Figure 6, are worse than withthe directional antennas. The values ofSgivenH0 haveincreased regardless ofT, to the range from 1.15 to 1.85.The values ofSgivenH1 now overlap with those givenH0 for T = 15s (and T up to 25 s), so regardless of thethreshold chosen, we cannot have a perfect detector.

We further believe that movement in the hallwayoutside of the room will have a greater impact whenusing dipole antennas, as compared to when using patchantennas. To test this, we have an experimenter stand

outside of the (closed) door waving his arms above hishead and moving from side to side. While this motion isoccurring, we run two more H0 experiments, one withdipole antennas, and one with patch antennas. Using thisdata, we recalculate the values ofSgivenH0. The resultsare shown in Figure 7.

The change in H0 values for Schanges slightly whenusing patch antennas, and changes significantly withdipole antennas. With patch antennas, the maximumvalue ofS given H0 has increased to 1.52 (comparedto 1.45 without motion outside of the door). With dipoleantennas, the max has increased to 2.08 (compared to

1 1.5 2 2.5 3 3.5 4 4.5 50

2.5

5

Prob.

Density


1 1.5 2 2.5 3 3.5 4 4.5 50

2.5

5

Prob.

Density


Empty

Person

Empty

Person

Fig. 6. Probability density functions for the S given H0(Empty) vs. given H1 (Person) for the (top) T = 15and (bottom) T = 30 second observation periods, for

experiments using dipole antennas.

1 1.5 2 2.5 3 3.5 4 4.5 50

2.5

5

Prob.

Density

Patch Antenna Network-wide Breathing Statistic S

Empty

Person

1 1.5 2 2.5 3 3.5 4 4.5 50

2.5

5

Prob.

Den

sity

Dipole Antenna Network-wide Breathing Statistic S

Empty

Person

Fig. 7. Probability density functions for the S whenT= 30s givenH0(Empty) with motion outside of the doorvs. given H1 (Person), for experiments using (top) patch

antennas and (bottom) dipole antennas.

1.80 without motion outside of the door). In this lattercase, even for medium T (T= 30s is shown in Figure7),there is overlap in the pdfs ofSgiven H0 and H1, andthus it is impossible to select a threshold for zero falsealarms and zero missed detections.

In conclusion, it remains possible to monitor breathingusing dipole antennas. However, it is clearly better,in terms of robustness to motion occurring outside ofthe room, to use directional antennas. Our experimentsusing patch antennas are only marginally affected by thismotion noise.


8/10


Actual Breathing Rate (bpm) 12.0 15.0 19.0Percent Estimates Invalid 0% 1.6% 0%RMSE of Valid Ests. (bpm) 0.08 0.42 0.30Bias of Valid Ests. (bpm) -0.03 -0.16 0.21

TABLE 1Rate estimation performance for three breathing rates

Person chest height (m) 0.88 1.01 1.13Percent Estimates Invalid 0% 1.6% 0%RMSE of Valid Ests. (bpm) 0.17 0.42 0.32Bias of Valid Ests. (bpm) 0.008 -0.16 0.24

TABLE 2Rate estimation performance for three bed heights

4.6 Rate Changes

In this section, we compare frequency estimation perfor-mance when the actual breathing rate changes. First, weperform three experiments with sensors using patch an-tennas in which the persons breathing rate is either 12.0,15.0, or 19.0 bpm. Using the metronome set at differentrates, the person ensures that their breathing follows thedesired breathing rate. We setT = 30s for all results. Theresults in Table 1 show best performance at the lowest

breathing rate tested, 12 bpm, but performance does notstrictly degrade with increased actual breathing rate.

4.7 Bed Height

In this section, we analyze three experiments with thebed (and thus the person) at different heights. In theexperiments discussed to this point, the persons chest

is at a height of 1.01 m above the ground. Here, we raiseor lower the bed height so that the persons chest is at0.88 m, 1.01m, and 1.13m, in three different experiments.At the lowest height, the sensors are nearly line-of-sight (LOS), that is the line connecting two sensors aremostly unobstructed by the persons body. At the highestheight, the sensors are predominantly at the height of themattress. In all experiments, the actual breathing rateis 15 bpm, and we use T = 30s. Table 2 shows the

breathing rate estimation performance. We find that thebest performance is in the nearly LOS case, about half theRMSE of the 2nd best height. Note that all bed heightsshow acceptable results, with a very small chance ofinvalid estimate, and RMSE below 0.5 bpm.

4.8 Fewer Sensors

In this section, we analyze what happens when we usea smaller number of sensors in the network. Since thenumber of links is proportional to the square of thenumber of nodes, we expect that having more sensorswill dramatically improve performance. In fact, the twosensor case is a limiting case which we have explored inSection4.1,which showed that one link is insufficient toreliably detect and monitor breathing.

7 10 13 16 190

0.3

0.6

0.9

1.2

Breathin

gRateError(bpm)

Number of Network Nodes

7 10 13 16 190

6

12

18

24

Percent

ofEstimatesInvalid

RMSE

Bias

Fig. 8. RMSE and bias of valid BreathTaking frequency

estimates for controlled breathing / patch antenna experi-ment (actual breathing rate of 15 bpm and person heightof 0.88 m), and percent of rate estimates that are invalid

(right y-axis), vs. number of sensors in the network T.Results are an average over 100 trials using randomly

selected subsets of the given size.

We use the same data collected with the 20-nodenetwork and test what would have happened with asmaller network as follows. Let Y {0, . . . , 19} be thesubset of nodes which we use in a particular trial. Werun tests with|Y |= 7, 10, 13, 16, 19. For each subset size,we run 100 trials and average the results. Subsets arerandomly selected in each trial from the full set of nodes{0, . . . , 19}.

Figure 8 presents the results for the RMSE, bias, andpercent valid of the breathing rate estimates. We can seethat estimates with just seven sensors in the networkare poor almost one in four breathing rate estimates isinvalid (greater than 5 bpm error). When the number ofsensors is increased to thirteen, the percentage of invalidestimates is 1.3%, and the RMSE of valid estimates is

below 0.3 bpm; and the results improve somewhat asthe network size increases to 16 and 19. Notably, thereare zero invalid frequency estimates with 19 nodes.

4.9 Phase Estimation

Beyond link amplitudes and network-wide frequencyestimates, there is also information to be gathered in

the phase of the sinusoidal signal due to the personsbreathing, in particular, for links which have a highamplitude |Al|. Consider that if two links measure asinusoid caused by one persons breathing, both should

be synchronous, that is, rise or fall at the same times.However, we do not know whether inhaling will in-crease or decrease the RSS on any particular link, so onelink may reach a maximum while another link reachesits minimum. In terms of phase, the {l}l might be radians apart from each other.

To show this effect in the data, we study the lestimates for T = 60s observation period. For example,


9/10


1

2

3

4

5

0o

90o

270o

180o

Fig. 9. Estimated phases, l for high-amplitude links l,shown on polar plots for five different experiments. Eachexperiments l values, to save space, are plotted on adifferent concentric circle labeled by experiment number {1, 2, 3, 4, 5}. Within an experiment, phases are seen tobe bimodal, with modes separated by 180o.

consider what we label experiment 1, a patch antennaexperiment with the person at height 1.01m breathingat 15 bpm. For this experiment, we plot l for linkswith the highest 5% of amplitudes Al in Figure9 in thecircle with radius 1. One may observe from the figurethat the phases are bimodal with two modes at about90 and 270 degrees. We repeat this plotting for fourother experiments, denoted experiments 2 through 5, ondifferent concentric circles on the same polar plot to savespace. Again, for each experiment, phase estimates forthe highest 5% amplitude links have two modes, in eachcase, separated by 180o.

These results are promising in the sense that whenthere are multiple stationary people breathing in thenetwork, we might be able to estimate the number ofpeople, even when the breathing rates are nearly identi-cal, by the number of modes in the distribution of phase.Further, we may be able to improve breathing detectionand monitoring using the fact that phase distribution is

bimodal. Such methods must be explored in future work.

5 RELATED WOR K

Breathing monitoring via capnography is standard prac-tice for anaesthetized patients in emergency departmentsand in intensive care units [12]. The capnometer usesan infrared absorption gas analyzer to measure thecarbon dioxide concentration in exhaled air (end-tidalC02). The breathing rate in this case is determined byfinding the frequency content of the CO2 concentrationsignal. The exhaled air is measured using tubes in thenostrils (e.g., nasal cannula), or from tubes connected toa face mask or tracheal tube, which are connected to the

capnometer. These tubes may become detached; if so, thecapnometer will detect apnea and alarm. Generally, themask or nasal cannula may be uncomfortable and limitthe patients movement. We propose a new non-invasivesensor (i.e., sensor not physically attached to the patient)for repiratory monitoring, which would allow a patientto sleep normally while being monitored. We note thatcapnography directly measures exhalation, while ourmethod indirectly measures breathing via the periodicchanges in the patients body size.

Breathing monitoring can also be performed us-ing plethysmography (respiratory inductive or thoracicimpedance plethysmography). These methods measure,using electrodes placed on the body, the change ininductance or impedance caused by inhalation and ex-halation. These electrodes can be contained in a bandworn around the chest. This is a method used in homemonitors for infants at risk of SIDS [13]. In comparison,the proposed system does not need to be attached to apersons body or have wires connected to the person.

Note that at physicians offices where proceduresrequiring sedation are performed, capnography andplethysmography are not typically used due to equip-ment costs. Instead, patients are monitored by a pulseoximeter, which measures oxygen saturation in blood. Ifa patient stops breathing, oxygen saturation decreases;however, the pulse oximeter will detect this desaturationonly minutes after breathing ceases [14].

Most closely related to the proposed system are otherproposed non-invasive breathing sensors. MicrowaveDoppler radar systems have been proposed for breathingrate estimation [15] [16]. Ultra-wideband (UWB) radarhas also been proposed for unobtrusive monitoring of

patients vital signs [17] [18]. In fact, UWB radars mayeven be sensitive enough to be able to detect a stationarypersons heart rate [19]. In comparison, our proposedsystem uses off-the-shelf wireless devices which aresignificantly lower in cost than radar devices. Ratherthan using a single high-capability radar transceiver, oursystem uses a network of many simple transceivers.

6 CONCLUSIONS

This paper presents a non-invasive respiration moni-toring technique called BreathTaking which uses signalstrength measurements between many pairs of wireless

devices to monitor breathing of an otherwise stationaryperson. We present a maximum likelihood estimator toestimate breathing parameters, including breathing rate,using all of the measured links RSS data simultane-ously. We present detection algorithms based on thoseestimated parameters, and an experimental testbed andprocedure to validate BreathTaking.

Using extensive experimental data collected with aperson lying in a hospital bed, we demonstrate theperformance of BreathTaking. We find breathing ratecan be estimated within 0.1 to 0.4 bpm error using 30seconds of measurements. We show that the links most


10/10


affected by breathing are the ones which receive lowaverage RSS. Breathing detection is demonstrated to re-liably distinguish between the breathing and its absenceusing 15 seconds of RSS data (in patch antenna experi-ments) and using 30 seconds of data (in dipole antennaexperiments), without false alarm or missed detection.Antenna directionality is useful to increase robustnessto external motion. Interestingly, the estimated phases oflinks which are affected by breathing distribution havea bimodal distribution with the two modes separated by180 degrees.

If BreathTaking is to be used for medical purposes,extensive evaluation on many people, and in manysettings, must be performed. In addition, this workexplored RSS-based breathing monitoring using 2.4GHz 802.15.4 radios the system may benefit fromtransceivers with other physical layer protocols andcenter frequencies. Finally, this sensor may be only onesensor used in a monitoring system, and its use withother sensors and sources of context information should

be explored.

REFERENCES

[1] H. Hill, C. Chapman, L. Saeger, R. Bjurstrom, M. Walter,R. Schoene, and M. Kippes, Steady-state infusions of opioidsin human. II. Concentration-effect relationships and therapeuticmargins,Pain, vol. 43, no. 1, pp. 6979, 1990.

[2] R. Litman, Conscious Sedation with Remifentanil During PainfulMedical Procedures* 1, Journal of pain and symptom management,vol. 19, no. 6, pp. 468471, 2000.

[3] F. Michahelles, R. Wicki, and B. Schiele, Less Contact: Heart-ratedetection without even touching the user, in Wearable Computers,2004. ISWC 2004. Eighth International Symposium on , vol. 1. IEEE,2004, pp. 47.

[4] T. Santiago and N. Edelman, Opioids and breathing,Journal ofApplied Physiology, vol. 59, no. 6, p. 1675, 1985.

[5] A. Vgontzas and A. Kales, Sleep and its disorders, 2003.[6] P. Sebel, M. Stoddart, R. E. Waldhorn, C. Waldman, and P. Whit-

field, Respiration: The Breath of Life. Torstar Books, 1985.[7] J. F. Murray,The Normal Lung: The Basis for Diagnosis and Treatment

of Pulmonary Disease, 2nd ed. W.B. Saunders Co., 1986.[8] J. Wilson and N. Patwari, See through walls: motion track-

ing using variance-based radio tomography networks, IEEETrans. Mobile Computing, vol. 10, no. 5, pp. 612621, May 2011,appeared online 23 September 2010.

[9] J. Wilson and N. Patwari, A fade level skew-Laplace signalstrength model for device-free localization with wireless net-works, IEEE Trans. Mobile Computing, appeared online 12 May2011.

[10] S. M. Kay, Fundamentals of Statistical Signal Processing: EstimationTheory. New Jersey: Prentice Hall, 1993.

[11] J. Wilson. SPIN: TinyOS code for RSS collection. [Online].Available: http://span.ece.utah.edu/spin

[12] T. Cook, N. Woodall, J. Harper, and J. Benger, Major compli-cations of airway management in the UK: Results of the fourthnational audit project of the Royal College of Anaesthetists andthe Difficult Airway Society. Part 2: Intensive care and emergencydepartments, British Journal of Anaesthesia, vol. 106, no. 5, pp.632642, 2011.

[13] M. Aurore Cote, C. Hum, M. Robert T. Brouillette, and M. The-mens, Frequency and timing of recurrent events in infants usinghome cardiorespiratory monitors, Journal of Pediatrics, vol. 132,no. 5, pp. 7839, May 1998.

[14] B. E. Hill, Development of an acoustic respiratory monitor,Ph.D. dissertation, University of Utah, Salt Lake City, Utah, USA,2011.

[15] M. Nowogrodzki, D. Mawhinney, and H. Milgazo, Non-invasivemicrowave instruments for the measurement of respiration andheart rates, NAECON 1984, pp. 958960, 1984.

[16] J. Whitney II and L. Solomon, Respiration rate signal extractionfrom heart rate, in Proceedings of SPIE, vol. 4368, 2001, p. 104.

[17] Y. Chen and P. Rapajic, Human respiration rate estimation usingultra-wideband distributed cognitive radar system, International

Journal of Automation and Computing, vol. 5, no. 4, pp. 325333,2008.

[18] S. Venkatesh, C. Anderson, N. Rivera, and R. Buehrer, Im-plementation and analysis of respiration-rate estimation usingimpulse-based UWB, inMilitary Communications Conference, 2005.

MILCOM 2005. IEEE . IEEE, 2006, pp. 33143320.

[19] N. Rivera, S. Venkatesh, C. Anderson, and R. Buehrer, Multi-target estimation of heart and respiration rates using ultra wide-

band sensors, in 14th European Signal Processing Conference, 2006.
http://span.ece.utah.edu/spinhttp://span.ece.utah.edu/spin

patwari11breathing.pdf

Documents