spirocall: measuring lun g function over a phone call · ngs. the spiro mulative volu le lung funct...

{

ACscPmothssthpsspWwa

AHp

AHM

IPmPuacto

PpnbcApaC©h

Spiro

1Computer SElect

UniverSea

{mayankg, saba

ABSTRACT Cost and accspirometers (dclinical settingPrior work, calmicrophone as or middle-incomhe latest smar

spirometry canstandard telephhe spirometry

printed vortex spirometry measystem, coineparticipants agaWe conclude twith or withouadvantages of e

Author KeywoHealth sensingprocessing; ma

ACM ClassificH.5.m. InformaMiscellaneous.

NTRODUCTIOPortability, lomobile phonesPhone-based husing and caadvantage is pchronic diseaseo monitor di

Permission to makpersonal or classronot made or distribbear this notice ancomponents of thiAbstracting with cpost on servers or tand/or a fee. RequeCHI'16, May 07-12© 2016 ACM. ISBhttp://dx.doi.org/10

oCall: MMayank G

E

Science and Etrical EngineDUB Group

rsity of Washattle, WA 98ae, emwhit, jwf

essibility havevices that ms, especially iled SpiroSmara spirometer.

me countries drtphones. In thn be performedhony voice cha

effort. We alswhistle can af

asures and mited SpiroCall,ainst two gold that SpiroCall ut a whistle foeach are discus

ords g; spirometry; chine learning

ation Keywordation interfaces

ON w-cost, and s a distinct a

health applicatiarrying dedicaparticularly appes, where patieisease progres

ke digital or hard oom use is grantedbuted for profit or cnd the full citatios work owned by

credit is permittedto redistribute to liest permissions fro2, 2016, San Jose, N 978-1-4503-336

0.1145/2858036.28

MeasuriGoel1, Elliot Eric C. Lars

Engineering,eering p hington 195 fromm, gaetano

ve impeded theasure lung fuin low-resourcrt, used a smart

However, inddo not typicallyhis paper, we d from any phannel to transmso investigate ffect the accurtigate usability , was evalustandard medihas an accept

or performing ssed.

mobile phone.

ds s and presentat

sensing capaadvantage in ions often savated medical parent in the ents frequentlyssion and ma

copies of all or pd without fee provcommercial advanon on the first pay others than ACM. To copy otherwists, requires priorom Permissions@aCA, USA 62-7/16/05…$15.0858401

ing LunSaba1, Mai

son3, Gaetan

, 2IHi

Kenmomwsti

o, shwetak}@u

he adoption function) outsice environmentphone’s built-

dividuals in lowy have access investigate ho

hone—using thmit the sound how using a 3racy of commochallenges. O

uated with 5ical spirometertable mean errspirometry, an

e sensing; sign

ion (e.g., HCI)

abilities provihealth sensin

ve patients frodevices. Th

management y use health tesanage treatmen

part of this work fvided that copies antage and that copiage. Copyrights f

M must be honorewise, or republish, r specific permissioacm.org.

00.

ng Funca Stiber2, Eno Borriello

Inglemoor igh School ore, WA [email protected]

uw.edu

of de ts. -in w- to

ow he of

3D on

Our 50 rs. ror nd

nal

):

de ng. om his of sts nt.

Recentdeveloheart ra[14], anSpiroSuser’s Spiromstandarsuch a(COPDparticiptheir luand cumultiplmanag

Introduapplicaaudio calculamodel sound step in

forareiesfored.toon

Figur(Sony

ction ovric Whitmir

o1, Shwetak N

028 m

3ComSo

tly, a numbeoped to estimate [13], resp

and newborn jaSmart, a smartp

lung function metry is the mard of care foras asthma, chD), and cystipants forcefullungs. The spiroumulative volule lung funct

ge various pulm

uced in 2012ation that recordata generated

ates the expiraof the vocal traround the use

n making spiro

re 1. A user y w580i) with, a

ver a Pre1, Josh FrN. Patel1

mputer Scienouthern Met

Dallas, eclarson@

er of health mate physiologpiratory rate [8aundice [5]. Laphone-based spusing the phon

ainstay for meadiagnosing ch

hronic obstrucic fibrosis. Dly exhale as mometer measurume of exhaletion measures

monary conditio

2, SpiroSmarrds the user’s d to a central atory flow ratract and a moder’s head. Spirometry more

using SpiroCaand without a S

Phone Comm1,

nce and Engithodist UniveTX 75205

@lyle.smu.edu

applications hgical measure8,12], pupillar

arson et al. [9]pirometer that mne’s built-in masuring lung fuhronic lung imctive pulmonaDuring spirommuch air as they

es the instantaned air. It then s to help diagons.

rt [9] is a sexhalation andserver. The s

te using a phydel of the reverboSmart was anaccessible, an

all on a featurSpiroCall whistl

Call

ineering ersity

have been s such as ry dilation introduced measures a

microphone. unction and

mpairments, ary disease

metry tests, y can from neous flow calculates

gnose and

smartphone d sends the server then ysiological

rberation of n important d since its

re phone le.

Medical Device Sensing #chi4good, CHI 2016, San Jose, CA, USA

5675

introduction, it has been involved in numerous clinical studies. SpiroSmart is currently deployed in multiple locations around the world, including Seattle and Tacoma in USA, Khulna in Bangladesh, and Pune in India. Thus far, we have collected data for around two thousand patients using SpiroSmart with encouraging results. While an analysis of the collected data is not the focus of this paper, we highlight four challenges that have surfaced from the SpiroSmart deployments: (1) SpiroSmart requires a smartphone; (2) usability and training challenges exist; (3) a patient with severely low lung function might not generate any sound; and (4) algorithms created from audio collected on a specific smartphone model may not generalize to other models or brands. In this paper, we critically examine ways to address these challenges and evaluate our proposed solutions with a set of 50 new patients.

Smartphones are becoming prevalent at a breathtaking rate, yet more than half of the mobile phone users in sub-Saharan Africa and South Asia will still be using a non-smartphone (or feature phone) in 2020 [4]. A major portion of the population suffering from lung impairments lives in these low resource environments. In fact, according to a recent WHO report, more than 90% COPD deaths occur in low- and middle-income countries [19]. Thus, we believe that phone-based spirometers need to work on all mobile phones, and not just programmable smartphones. Even smartphones, the diversity of phone manufacturers and models makes it challenging to manage custom applications for every type of mobile phone.

To this end, we present SpiroCall (Figure 1), a call-in service that measures lung function on any mobile phone without the need for a locally running application. Unlike SpiroSmart, it transmits the collected audio using the standard voice telephony channel. A server receives the data of degraded audio quality and calculates clinically relevant lung function measures and reports to the participants using audio or text message. The ability to use a server to analyze audio data transmitted from any mobile phone, be it a feature phone or smartphone, eliminates the need to develop a specialized application for every phone platform. SpiroCall combines multiple regression algorithms to provide reliable lung function estimates despite the degraded audio quality over a voice communication channel.

Although the call-in service removes the need for a smartphone, there are other significant usability challenges that are more difficult to mitigate: how a user holds the phone (angle, microphone occlusion, etc.), the distance from the user’s mouth to the phone, and how wide a user opens their mouth. Recognizing that some people may not be able to master the technique needed to perform this maneuver, we also designed a simple and low-cost 3D-printed whistle accessory. The whistle (Figure 1, Top) generates vortices as the user exhales through it [17,18], changing its resonating pitch in proportion to the flow rate.

The whistle does not have any moving parts and is as simple as any spirometer mouthpiece. Despite the additional hardware, the whistle offers several important advantages: (1) the acoustic properties of the whistle are more consistent than a user’s vocal tract and generate audible sounds even at lower flow rates, (2) the whistle removes the effect of distance from the user’s mouth, and (3) precisely controlling mouth shape and phone orientation are less important. In this paper, we investigate viability of the call-in service approach with and without the whistle.

We evaluated SpiroCall in a controlled study with 50 patients. We compare SpiroCall to two FDA approved spirometers and evaluate the effect of using the voice communication channel on the performance of SpiroCall. Each patient performed spirometry efforts with and without the whistle on two different phones recording the audio through the cell phone network and two smartphones recording the audio locally through an app. Participants used two different sizes of vortex whistles to determine whether different sizes work better for different individuals. Our results show that without a whistle, SpiroCall has a mean error of 7.2% for the four major clinically relevant lung function measures. For FEV1% (the most commonly used diagnostic measure [2]), the mean error is 6.2%. With a whistle, SpiroCall has a mean error of 8.3% for the four measures, and 7.3% for FEV1%. Although, using the whistle leads to higher average error in lung function estimation, it performs more consistently for people with lower lung function and produces fewer over-estimations of lung function (i.e., false negatives), as compared to when not using a whistle.

The main contribution of this paper is a demonstration that every mobile phone in the world can be used as a spirometer. This contribution comes in four parts: (1) an algorithm to estimate lung function from a standard telephony voice channel’s degraded audio signal; (2) a custom-designed whistle that reduces usability and performance challenges; (3) a comparison of the call-in service and the whistle against two clinical spirometers (using different phones); and (4) a demonstration of how poor quality audio, transmitted across the standard telephony voice channel, can be utilized for modeling and inference.

BACKGROUND OF SPIROMETRY Spirometry is the most widely employed pulmonary function test. Many different types of spirometers are available, ranging from big, clinical spirometers to portable, home spirometers. Their cost also varies from $1,000 USD to $5,000 USD. During a spirometry test, the patient takes the deepest breath possible and then exhales with maximum force for as long as possible. The spirometer measures the amount and speed of airflow and calculates various lung function measures based on the test. Four of the most important lung function measures are:


5676

(

(

(

(

Aah(

•

•

•

Agtihaof((fr(

DTcvlovaluldcecc(aPp

Fm

(1) Forced Vitexpelled dur

(2) Forced ExVolume of a

(3) FEV1/FVC

(4) Peak Expiflow rate rea

A healthy indivat least 80% oheight, and ge(expressed as a

Mild to Me

Moderate L

Severe Lun

Apart from thgenerate flow ime plots. Fig

healthy individalmost a straigobstruction to tfaster than exp(PEF). Therefo(blue, dashed lfrom a restrictirespiratory mus(FVC) decrease

DESIGN OF SPThe previous wchunk of compvia an Internetossless and fre

voice communa server. The seung function mearning model

detection, speccoding (LPC). encode voice. channel comprcoefficients rem(Plain Old Telea communicatiPOTS in this pprevalent than l

Figure 2. Exammajor lung func

tal Capacity ring the expirat

xpiratory Voluair expelled in t

C (FEV1%): R

iratory Flow ached during th

vidual’s lung fuof the values ender [7]. Abna percent of pre

edium Lung Dy

Lung Dysfunct

ng Dysfunction

he numerical vs. time, flow

gure 2 shows dual, the desceght line (blackthe airflow inconentially afte

ore, it attains line in Figure 2ive lung diseascles weaken aes (red, dashed

PIROCALL work of Spiroutation to a sert connection. Tee of artifacts. ication channeerver then usesmeasures. The l fall into threectrogram procThe cellphoneThis means

resses the audmain largely pephone Serviceion channel. Hpaper becauselandlines in the

mple of differenction measures

(FVC): Totaltion,

ume in one sthe first second

atio of FEV1 an

(PEF): Maximhe test.

unction measurpredicted base

normal values edicted value) [

ysfunction: 60-

ion: 40-59%, a

n: below 40%.

measures, sw vs. volume,

examples of ending limb ofk, solid line increases, the flower reaching its a curved or “

2). For an indise, such as cyand the patient

d line in Figure

oSmart offloadrver, with the aThus, the receIn SpiroCall,

el to transmit ts machine learfeatures used categories: tem

cessing, and le channel (GSM

that even thdio signal, the preserved. Lande) is also an attHowever, we d GSM networe developing w

nt Flow vs. Vols.

l volume of a

second (FEVd of expiration

nd FVC, and

mum expirato

res are generaled on their ag

of FEV1% a[10]:

-79%,

and

pirometers aland volume vFV plots. In f the FV plot n Figure 2). Aw rate decreasmaximum valu“scooped” slopividual sufferin

ystic fibrosis, tht’s lung capaci2).

ded a significaaudio transferreived audio wwe leverage th

the audio data rning to compu

by our machinmporal enveloplinear predictivM) uses LPC ough the GSvalues of LP

dlines, or POTtractive option do not focus orks are far moworld.

ume curves an

air

1): n,

ory

lly ge, are

so vs.

a is

As ses ue pe ng he ity

ant ed

was he to

ute ne pe ve to M

PC TS as on

ore

Lung FFigure recordeafter th(Right)There spectrogoes thcommutechniqshare supon th

First, amodel or glottract blof the channemethodvocal calculapreservcoefficfeatureencodithe sigfundam(insidehigher spectra

AdditioenergyFigurecuts offidelityrelativein Figu

FigurelocallyGSM 4 kHz (red sq

nd

Function Estim3 shows sp

ed locally on ahe data is sen). The device

are significograms as the hrough a communication netwques, all GSsimilarities in the same underl

all GSM voice for speech. T

ttis excitation, alurs this excitasource and f

el, instead of d for separatintract filter is

ation is that ved (and arecients). Thesees in our algoing to preservegnal. An exammental resonane the red box),

harmonics oal energy above

onally, the try components 3 (Right) wheff in patches.y and stays abely louder) segure 3, Right).

e 3. (Left) Spey, and (Right) rnetwork downis lost. Howev

quare) is large

mate without Wpectrograms a smartphone (nt through a G

type is an iPcant differenc

voice undergmunication chaworks use d

SM/UMTS sptheir treatmentlying linear pre

coding technoThat is, the “so

and the “filter”ation into contifilter are then f the raw audng out the souto use LPC. the strong fre

e calculable de resonances orithms. As sue much of the

mple of this is nce is easily , despite manyf the fundame 4 kHz, being

ransmission pin the signa

ere the energyIn contrast, th

bove the noise gment of the e

ectrogram of a recorded throunsamples the ver, the data in ly reconstructa

Whistle of a spirome(Left) and the s

GSM cell phonPhone 4S in bces between goes many chaannel. Althougdifferent speecpeech-codingt of speech andediction appro

ologies use a soource” estimate” estimates howinuous sound. P

transmitted thdio. The mosurce excitationAn artifact oequency resondirectly from

are also theuch, we expecimportant infoshown in Figuseen in both

y smaller detaimental resonancg lost.

process suppreal, as can by of the signahe signal mainfloor for the i

effort (inside th

spirometry effogh the voice caudio, and theour main regio

able.

etry effort same effort ne network both cases.

the two anges as it gh different ch coding algorithms

d are based ach.

ource-filter es the lung w the vocal Parameters hrough the t common n from the f the LPC nances are

the LPC e primary

ct the LPC ormation in ure 3 – the recordings

ils, such as ce and all

esses low-e seen in al abruptly ntains high initial (and he red box

ort recordedhannel. Thee data over

on of interest


5677

ACIaep

Wathrpabarrththabth““eaEe

WfPTvFingwPee

Csrfrpbcthein

Itoulip

Algorithm for Channel withouIn order to deaas the data goevaluate what mproposed in the

We use the miand the receivhree approach

resonance trackpredictive codinassumed to be because it is aamplitude) at resonances canreflections in herefore shoulhem. Finally,

approximation.be divided into he source po

“filter” in our“source variancexciting the approximates tEach approachestimations.

We extract sepfrom these timePEF is definedThus, for a givalue and use iFVC is definedntegrating the

gives us a featuwe generate 3 PEF, each. Westimate FEV1estimated FEV

Considering thsound, the LPremain largelyfeatures are barespect to timpreserved as wby the GSM chcompletely loshe resonances

enough that mnformation.

In the original o a random

underlying lineinearity in th

poorly on the d

Lung Functiout a Vortex Whial with the draoes through a modifications e original Spiro

crophone as anved pressure vhes (Figure 4)king in the freng (LPC). Thea reasonable a

a measure of low frequenci

n be assumed the vocal tra

ld be proportiowe can use

Linear predica source and a

ower and shar case approxce” is an estim

vocal tract the power of th generates m

parate feature se-domain flow-

d as the maximven flow-rate it as a feature fd as the total

e flow-rate esture for FVC resets of of 38

We do not use 1%. This valu1 and FVC.

hat the GSM PC-based featuy preserved. Tased on the come; in most well. The specthannel becauset. However, us within the

most spectral fe

algorithm, theforest regress

ear model, had he feature datdata collected

on Estimation istle

astic variation GSM channe

are necessary oSmart paper [9

n uncalibratedvalues are tra): (1) envelopquency domain

e envelope of tapproximation the overall si

ies. In the freto be amplitu

act and mouthnal to the flowlinear predic

ction assumes ta shaping filteraping filter coximates the v

mate of the whifilter—in o

the flow rate ultiple time-do

sets for FEV1,-rate estimation

mum flow reachestimation, w

for PEF regressvolume of air

timation with egression. Usinfeatures for Fany regressio

ue is simply

channel uses ures used in The envelope oarse amplitud

cases these tral features are the high frequupon analysis w

first harmonieatures contain

e calculated fesion that, bectrouble exploi

ta. The algorthrough the G

over a Voi

in sound qualiel, we sought to the algorith9].

d pressure sensansformed usinpe detection, (n, and (3) linethe signal can bof the flow ra

ignal power (quency domaiudes excited bh opening—an

w rate that causction as a flothat a signal cr and it estimatoefficients. Th

vocal tract. Thite noise proceour case, thfrom the lungomain flow-ra

, FVC, and PEns. For examplhed in an effo

we take the masion. In contrar exhaled. Thurespect to tim

ng this approacFEV1, FVC, anon algorithm

a ratio of th

LPC to encoour algorithmdetection-bas

e of sound wifeatures remare most affectuency details awe realized thics were stronn some releva

atures were secause it has niting some of thithm performSM channel an

ce

ity to

hm

sor ng (2) ear be ate (or in, by nd ses ow an tes he he

ess his gs. ate

EF le,

ort. ax st, us, me ch, nd to he

de ms ed ith ain ed

are hat ng ant

ent no he ed nd

it overobstrucupdatediffereeach rerun theand uoverfitthe algfor the

The firlinear rlung furegressfeatureunderlyregresseliminaThis reand ridstable. regressdata inThoughfitting find ththese fFVC, calculavalues.

Additioin a noTo dedetectsmean 250 msis the popportuamplituabove environThe thSpiroS

Figurewhistle

r-estimated thcted lungs (F

ed the algorithent regression aegression woue regressions u

use leave-one-tting. Furthermgorithms at thee collected data

rst regression irelationship beunction value. sion (LARS) es using a variaying model ision uses theates features inegression usesdge regression

Finally we sion (k = 2) [6n the feature sph the underlyioften can fit m

he final regresfour regression

and PEF mated as a rati.

onally, there aroisy environmeeal with such s the level of absolute amp

s window immperiod when thtunity to measude of sound

an empirinment is conshreshold used Smart clinical tr

re 4. Flowchae.

he lung functioFEV1% < 0.8hm to employalgorithms (Figuld provide a dusing the sciki-patient-out cr

more, we keep eir default valua.

is a linear regretween the featThe second re[3]. LARS s

ant of forward is assumed toe elastic net n a slightly dis a combination for regulariz

use enclosin6], which findpace and fits a ing model is a

many different tsion estimate

ns. We use thismeasures. As

io of the esti

re situations went or the chan

situations, thf background nplitude of themediately beforhe user is most sure the ambwithin this wiically determsidered unsuitahere is same

rials.

rt for lung fu

on for particip8). Therefore,y an ensemblgure 4), with thdifferent perspit-learn toolkitross-validationall the parame

ues and do not

ression that trietures and the gregression uses selects the mfeature selecti

o be linear. algorithm [2

ifferent way thon of LASSO zation that is ong k-Nearest

ds the convex locally linear

assumed lineartypes of nonlinby taking the

s same process mentioned, F

imated FEV1

when the test is nnel itself mighhe system autnoise by look

e recorded sore the user exhsilent and we u

bient noise levindow is estimmined threshable for data as the one u

nction estimat

pants with we have le of four he aim that ective. We t in Python

n to avoid eters for all t tune them

es to find a round truth least angle

most useful ion, but the The third

20], which han LARS.

regression often more

Neighbor hull of the regression. r, the local nearity. We median of for FEV1,

FEV1% is and FVC

performed ht be noisy. tomatically king at the und for a hales. This use it as an vel. If the

mated to be hold, the collection.

used in the

tion without


5678

LSvsdBcvmspetordpmtaHsp

TcccthinaacmFwscarg

Ts[

wosdtuobithv

Lung FunctionSpiroCall facevariability of dseverely impaidistance betweBernard Vonnchanged its pitvortex whistle modifications spirometry effopitch tracking estimate of flowo variations a

response, the wdistance, channparts and thus, mass-produciblest the design

a flow-sound tHowever, we spirometry andparticipants.

The vortex whcylindrical cavcylindrical pipcylindrical cavhrough this tunside to swirl a

attached perpenair enters the ccircumference moves towardFigure 5 (Left) within the whsimulated velocylindrical cavaround at an arotational velogenerates sound

The frequencyseveral factors17]:

where is freqof the vortex sectional area downstream tuube, and Δ

outlet of the between the fors difficult to he quantity be

values range be

n Estimate wites three aud

different phoneired patients, een a user’s negut, in 195tch in proportio

[17]. Later, Wto vortex wh

forts [16,18,2to convert th

w rate. Considacross devices,whistle could mnel, and deviceis as simple asle for less thanproposed in [1transducer instfound the pro

d modified it b

istle consists oity, and the dope that is tavity on its curube. The cyliaround the chandicular to thecylindrical cavof the cavity,

d the downstshow the resu

histle. The colocity of the avity, the vortexangular velociocity of the vd as it leaves th

y produced bys, including th

quency, is in

in the cylindof the inlet,

ube, is th is the lengt

whistle. The rmed vortex ancalculate mathdetermined th

etween 0.35 up

h Whistle dio sensing s, (2) low sounand (3) inconmouth and t

54, designed on to flow rateWatanabe and histle construc21]. In their se vortex whisering pitch trac, such as gain

make SpiroCale. The whistles any spirometen 10 cents (US)16] to see if it ctead of the usoposed design

based on a pilo

of three sectionownstream tubangentially corved surface. ndrical cavity

amber. The dowe cylindrical c

vity, it starts rothereby formintream tube. ult of a simullor of the arrair. When thex becomes unsity that is provortex. This he downstream

y the whistle he dimensions

nput flow rate, drical cavity,

is radius ohe length of th of the vortesin term refend cylindrical hematically, nrough calibrati

p to 0.95 [16].

challenges, (nd amplitude fnsistency of ththe microphon

a whistle the and called it

Sato suggesttion for use tudy, they us

stle sound to cking is resilien and frequenl independent

e has no moviner mouthpiece—). We decided could be used

ser’s vocal tracn unsuitable fot study with

ns: the inlet, thbe. The inlet isonnected to thThe user blow

y allows the awnstream tube avity. When th

otating along thng a vortex, anThe arrows lation of airflorow denotes the air leaves thstable and whioportional to th

unstable vortm tube.

is affected bs of the whist

(1) is the radiis the crosof the air in ththe downstrea

ex formed at thers to the angplane. This terecessitating thion [16]. Typic

(1) for he ne. hat t a ed in ed an

ent cy of ng —to as ct. for 15

he s a he ws air is

he he nd in

ow he he ps he ex

by tle

ius ss-he

am he

gle rm hat cal

We mowhistle15 L/srate isindividsizes oas diffe

We 3Dusing whistlebigger pitch gparticipit was

AlgorithWhen proceslinearlyestimatparamea few whistlebe calib

Whistl44.1 kdevicespectrodata inbetweespectrowithin the wheffort. that potowardat the

Figurshow fastertwo wLDST == 74,

odified the dese’s response res (verified via s well aboveduals with heiof the whistle (ferent sizes will

D-printed the wABS plastic

e sizes with 5whistle perfo

gradient. Frompants also prefeasier to handl

hm for Lung Fua vortex whissing consideray in response tte the flow reters of this linexample spiro

e with set dimibrated once.

le Pitch ExtrakHz to ensures with differenogram of the efnto frame duraten frames. Nexogram (Figure

0.25 seconds.hite curve) cor

We track pitcoint, stopping wds lower frequend of a spiro

re 5. (Left) 3Dthe airflow an

r and blue is slwhistles: Small:= 24, LCC = 40.LDST = 27, LCC

sign suggested emains linear eSolidWorksTM

e the peak fight up to 210(dimensions shl have differen

whistles on a Smaterial. Our 50 participantormed better bm a usability sferred using thle.

unction Estimastle is used, wably because thto flow rate. S

rate over timenear relationshometry efforts

mensions, these

action: All aue uniformity innt sampling raffort to track thtions of 46 msxt, we find th6) and search

. The peak frerresponds to thch backward anwhen the spectruencies. This hometry effort t

D rendering ofnd the colors dlower. (Right) : RCC = 20, RD

Big: RCC = 37= 35.

in [16] to ensueven at flow ra

M simulations).flow rate atta0 cm. We deshown in Figurent pitch gradien

Stratasys BST7evaluation o

ts demonstratebecause it hadstandpoint, 34 e bigger whistl

tion with Whiste can simplifyhe whistle pitc

Simple pitch tre. We can cahip (bias and sl. For a particuparameters on

udio data is resn the process

ates. We first phe pitch. We ses with a step sie peak magnith for the peak quency (Figur

he PEF of the nd forward in ral energy ceas

helps us ignorethat may over

f the whistle. denote the veloDimensions (in

DST = 11, RIT =.5, RDST = 12.5

ure that the ates around This flow

ainable by signed two e 5, Right), nts.

768 printer f both the

ed that the d a steeper

out of 50 le, because

tle y the audio ch changes racking can alibrate the lope) using ular vortex nly need to

sampled to sing across process the egment the ize of 3 ms tude in the

k frequency re 6, top of spirometry time from

ses to trend e wheezing rwhelm the

The arrowsocity. Red isn mm) of the= 8, LIT = 50,5, RIT = 8, LIT


5679

mqtwfrpw

FTp

T(ininweesc

Figure 7. Flwhistle.

main whistle aquadratic polynwo neighbors

frequency estimresonance passpitch thresholdwell at lower fl

For example, inThis means thapitch data, the F

Tail Extrapola(PEF), the flondividual andndividual with

when we extrapexponential fitexponential andsystem automacurves, includin

Figure 6. Thewith the pitch textrapolated b

owchart for l

audio amplitudnomial to the frs to attain sumates [11]. Wses below a c, as the whistlelow rates and th

n Figure 6, the at while we canFVC value nee

ation: After thw rate decays

d decays fasteh obstructive polate the pitcht function. Wd exponential tically adapts tng the ones wh

e blue and ortracked by the

based on inform

ung function

de. For each frequency bin oub-bin accurace stop pitch tr

certain empirice mouthpiece dherefore lower

pitch can be trn infer the FEVeds an extrapol

he flow achieves exponentialler than exponlung impairmh curve, we ca

We apply a of exponentialto different typhere the flow r

range regions algorithm. The

mation in the bl

estimation w

frame, we fit of interest and icy in our peracking once thcally determindoes not resonar frequencies.

racked up to 1V1 value from th

ated curve.

es its peak valuy for a healthnentially for

ments. Thereforannot just use combination l fits, so that thpes of flow-timrate decay fast

are associatee green region ue region.

with

a its ak he ed ate

s. he

ue hy an re, an of he

me ter

than exend of

We useour extFigure extrapophase iis avairesonanresonanregion.applyinensure exhalatexceptiwhen receiveaveragFVC tcurve a

FVC Rmethodmeasurmost cas a sfeaturedown iHead, peak frfeaturepeak amBody strackinarea uestimatextrapoextrapoof the Equatiofeaturedemogto recohelps functio

SimilaralgorithLARS,of all rthe finvalidat

edis

xponentially . Wf the flow-time

e the entire destrapolation fun

6 shows tholated. The bluin time during ilable. Howevence-tracking tnce-tracked da. We evaluang it to the se

it was able tion. Although

tionally on growe evaluated

ed from SpiroCge error of 15%through a regras a feature in t

Regression Md did not pre, it still prov

cases. We, therset of regressies used in the rinto the three pBody, and Ta

frequency of the from the Heaamplitude (normsection of the cng curve until under the curvte. The next olation. We usolation as an e curve. Specion (2) as featu

es, we also graphic featureord a patient’the device in

on for the patie

r to the nohm employs an, and elastic neregressions an

nal FVC estimation in all level

We fit the follocurve:

scending limb nction. The grehe time oveue area (Body,which reliable

er in order to tto extrapolatioata to extrapolaated our extret of groundtr

to model thh the extrapooundtruth data

the extrapolaCall devices, o

%. We thereforression modelthe regression.

Model: Althougprovide an advided a good, refore, encode ion features. regressions. Thphases of the pil. We use thehe tracked pit

ad section of thmalized) of thecurve, we use the end of th

ve until the eset of feature

se the coefficieencoding of theifically, we uures in the regr

use height, s. It is commos physical detn calculating ent.

o-whistle conn ensemble of et regressions. d select a medate. We use leals of learning t

owing function⋅ of the tracked

een area (Tail Fer which the Data+Fit) repe resonance tratransition smoon, we cross-ated data withrapolation furuth flow-timehe tail end oolation functioa (mean erroration on the a

our FVC estimare decided to el, using the ex.

gh our tail exdequate volumalbeit noisy, ethe pitch trackFigure 7 show

he features canpitch tracking ine estimated PEtch, as the rephe curve. We ae overall audiothe area under

he body sectioend of 1 sec, es comes froments generated e curve in the Tuse , , anression. Apart age, and se

on practice in stails as this inpredicted no

ndition, the f three regressio

We combine tdian of their esave-one-patieno avoid overfit

n to the tail (2)

pitch to fit Fit only) in

curve is presents the acking data othly from -fade from

hin the blue unction by e curves to f a user’s on worked r = 3.2%), audio data ates had an stimate the xtrapolated

xtrapolation me (FVC) estimate in king output ws all the

n be broken n Figure 6:

EF, i.e., the presentative also use the o. From the r the pitch-

on, and the i.e., FEV1

m the Tail by the tail

Tail region nd from from these

ex as our spirometers nformation

ormal lung

regression ons: linear, the outputs stimates as

nt-out cross tting.


5680

ETaramcam

Tf

•

•

•

Ar

EWsEfTthlodpam

Figure 8. Spdata on four pthe audio locanumbers and the Google Vo

EVALUATION To evaluate Spaudio samplesrecruited 50 paage from 21 tomessages in conducted in approximately moderate lung

The SpiroCallfactorial design

Phone TypeiPhone deviW580i. We the performyear-old dev

Channel Trecording. channels to the channel

Whistle: Norecorded audifferent pasize gave re

All the conrandomized the

Experimental SWe collected thsmartphones, aEricsson W580front of the useThe distance whe iPhones andocally at 32 kH

devices sent thphones placedaccounts that rmessages. Goo

iroCall experimhones at the s

ally. The other tsent the audio

oice server.

piroCall, we crs and groundarticipants (30

67 years (M =the universitya non-clinica30 minutes. 20obstruction, i.e

l study used n. The factors a

e: iPhone and nices: Samsungused the W58

mance of Spirovice.

Type: Local We kept theanalyze the peis changed.

o whistle, smaludio data for tarticipants prefesults that were

nditions weree order of the w

Setup he audio data oa Samsung G0i feature phoer at roughly anwas not formalld the SamsungHz and 44.1 kHhe data over thd phone callsrecorded the d

ogle Voice sav

mental setup. ame time. Twotwo phones cao data over the

reated an exted truth spirommales, 20 fem

= 30) through y. The study al lab setting 0% of participe., FEV1% < 0

a within-suand levels were

non-iPhone. Wg Note 3 and80i (feature phCall on an ap

recording or e iPhone conerformance of S

ll whistle, and two whistles tferred differente more reliable

counterbalawhistles.

on four phonesGalaxy Note one. All four n arm’s length ly controlled og Note recordeHz, respectivelyhe GSM voice to different

data in the fored the audio d

We recorded o phones recoralled Google Voe GSM channe

ensive dataset metry data. W

males), ranging flyers and ema

sessions weand lasted f

ants had mild .80 (Table 1).

ubjects 2×2×e:

We used two nod Sony Ericssohone) to evaluapproximately 1

voice channnsistent in boSpiroCall if on

big whistle. Wto understand t sizes or if onthat the other.

anced and w

s, two iPhone 43, and a Sonphones were away (Figure 8

or varied. One ed the audio day. The other tw

e channel. TheGoogle Voi

rm of voicemadata as 44.1 kH

thededoiceel to

of We

in ail ere for to

×3

on-on ate 0-

nel oth nly

We if

ne

we

4S ny in

8). of

ata wo ese ce ail Hz

MP3 fless trecordiin Figu

We traand Saat the Google

ProcedWe coFDA-aLegendtwo spparticiphow mdevicedeviceperformspiromtwo wmeasurfatigueTherefthe begthe endparticip

At theexpiratto pracable toATS crecordeparticip

The fothe audperformthe aueffortsaccepta

Table

Males

Age (y

Heigh

Repor

Asthm

Cystic

Low L

Never

files, but the Gthan 8 kHz. ings and thoseure 3.

ansferred the damsung Note) tend of the stue Voice accoun

dure ollected the groapproved clind and the NDDpirometers to apants got fatig

much variabilitys. We recordes to use it mance. The

metry efforts (thwhistles, and rements are ce can build upfore, we recordginning of thed of the sessiopant.

e start of eactory maneuverctice using the o perform an criteria for reped using the pants to SpiroC

our phones (Phdio simultaneoming tests wituthors, who w, gave feedbaability and qua

e 1. Demogra

Participa

s (n, %)

yrs) (mean, ra

ht (cm) (mean,

rted Lung Ailm

ma: 10 (20%), B

c Fibrosis: 1 (2

Lung Function

r Performed S

GSM channel The differen

e done over the

data from the to the compute

udy. We downlnts as MP3 file

ound truth fornical spirometD EasyWare sanswer two qugued as the sesy exists betweeed the variabil

as a bencparticipants p

hree each for: one withou

completely effp when perforded efforts on oe session and oon. We random

ch session, wr to the participspirometer. Oacceptable maproducibility spirometer. N

Call.

hone Type × ously, thus savth each devicewas trained toack to the paality of the eff

phic informat

ant Demograp

ange)

, range)

ments

Bronchitis: 2 (4

%), Sarcoidisi

n (n, %)

Spirometry (n,

band-limited tnce between e GSM channe

local phones er over a USB loaded the dates to a compute

r the participanters: the nSpspirometer. Wuestions: (1) wssion progresseen the outputslity between t

chmark for Sperformed at two clinical sp

ut whistle). Sfort-dependent rming this maone clinical spon another spimized the orde

we explained tpants and we ance the partici

aneuver accord[10], three ef

Next, we intro

Channel Typeving the partici type separateo administer articipants regforts. In the fut

tion of the par

phics (N = 50)

30 (60%

30 (21 –

172 (15

4%), COPD: 2

s: 1 (2%)

16 (32%

, %) 30 (60%

the data to the local

el is shown

(iPhone 4S connection ta from the er.

nts on two pire Koko

We used the whether the ed, and (2) of the two

the clinical SpiroCall’s

least 15 pirometers, Spirometry and some

any efforts. irometer at irometer at er for each

the forced asked them ipants were ding to the fforts were oduced the

e) recorded pants from

ely. One of spirometry

garding the ture, it will

rticipants.

%)

– 67)

55 – 188)

(4%),

%)

%)


5681

bd

NsgScs[

RIwapaSdawaSSw

TAgmoFrsaptwowwthesrwd

T

be straightforwdetermines if an

Note that collspirometer dataground truth SpiroCall was aclinical spiromselects the eff2,10].

RESULTS In this sectionwhen comparedaccuracy of espositives vs. faa false negativeSpiroCall predidown these resalso compare without a vorteand usefulnessSpiroCall. BasSpiroCall can with lung impa

Two Ground TAs mentioned, groundtruth. Wmeasure and foof 9.2% betweFEV1% had respectively. Hstatistically sigalso studied thplayed any rolewo devices. I

order of the twwe found that twere statisticallhat the partici

ended. Therefspirometer thatreference devicwas not considevice perform

Figure 9. Perc(Right). The firThe error bars

ward to have n effort was to

ecting the Spa at the same was unknownassociated with

meter. As per thfort with the h

n, we discuss d to the two cstimated lung

alse negatives. e if the groundicts the value tsults by Phone

the performaex whistle. Fins of the flow-sed on our ehelp in scree

airments in low

Truth Devices we used two

We compared tound that PEFeen the two da difference However, nongnificant (basehe effect of oe in exaggeratin a 2-way AN

wo spirometersthe difference ly significant (pipants got fatifore, we use t the participance. While this mstent across p

mance was not

cent error for dist two devices show standard

a system thao low in volum

piroCall data atime is impossn. Instead, eah the best efforhe ATS criteriahighest FVC

the performanlinical spiromefunction mea

We consider adtruth FEV1% to be above 0.

e Type and Chance of Spironally, we discu-volume curveevaluation weening and mon

w resource regio

clinical spiromtheir respectiv

F had the maxidevices, and F

of 5.1%, 5.2ne of these ed on an F-tesorder to undering the differeNOVA test ws as a betweenin estimates of(p<0.05). This igued by the t

the results nts used as theimeans that the participants, tht found to be

fferent lung funrecorded the d

d deviation.

at automaticalme.

and the clinicsible, so explicach effort frort selected by tha, the spirometas the best on

nce of SpiroCaeters in terms asures and falan estimate to bis below 0.8 an.8 [2]. We bre

hannel Type. WoCall with anuss the accuraces generated be conclude thnitoring patienons.

meters to colleve lung functioimum differenEV1, FVC, an2%, and 3.2%differences a

st, p>0.05). Wrstand if fatigunce between th

with presentation-subjects factof PEF and FEVfinding suggestime the sessiofrom the fir

ir groundtruth reference devi

he difference significant an

nction measuredata locally in a

lly

cal cit

om he ter ne

all of

lse be nd ak

We nd cy by hat nts

ect on

nce nd %, are We ue he on or, V1 sts on rst or ce in nd

shouldcorrectPhone,particip

Lung FWe brSpiroCwell itthe num

AccuraThe greach mmeasurthan 1performapp (Sover thApple whichin diagThe meconditifunctio[15]. Fexpecta smarHowevhigher the lunestimatfirst stcriteria

OutlierIn ordemodifishowsoutput of FEVsolely functiois positnegativRight,

es on different dan app; the nex

d not strongly ated for fatigu, Channel, anpants.

Function Estimreak down the

Call and the clt performs for mber of outlier

acy of Lung Funraphs in Figuremeasure withores, the algori0%. There is mance of smaramsung Note ahe voice commiPhone 4S). Tis the most co

gnosis becauseean error rate fions. The ATon measures toFor most patiented level of varartphone and pver, it is impor

than twice thng function mted. We use twtandard deviata [15] and the r

rs and Patientser to understanied Bland-Altmthe percentageof a spiromet

V1%. Lines inon FEV1% b

on measure fortive, then the lve). It can be that SpiroCal

devices withouxt two devices

affect the final ue by counternd Whistle T

mate without We comparison linical spirom

r different lungrs and false neg

nction Measuree 9 (Left) presenout a whistleithm returns ano significant

rtphones recordand Apple iPhmunication chThe performanommon measue it is typicallyfor FEV1% is b

TS acceptabilio be within 7%nts, SpiroCall riation, even ifperformed thertant to evaluathe standard devmeasures are uwice the standation is withinresult cannot b

s with Low Lungnd the directionman plots [1] ie difference beter versus the ndicating ±2σ because it is r diagnosis. If tlung function wseen in Figurl (both withou

ut whistle (Left)recorded the d

analysis. In adr-balancing beType condition

Whistle of measurem

eters by evalug function megatives.

es nt the percenta. For all lungan average errt difference beding the data lohone) and phonhannel (Sony Wnce is best fo

ure of lung funy most consistebelow 6% for ity criteria req% to 10% of operforms well f the patient die test on a pte the outliersviation) and seunder-estimatedard deviation bn the ATS ace considered a

g Function n of the bias, in Figure 10. etween SpiroCspirometer me(red dashes).

the most comthe percentagewas over-estimre 10, Top-Lefut whistle) ten

), and with whisata over a pho

ddition, we etween all ns for all

ments from uating how asures and

age error of g function ror of less etween the ocally in an nes running W580i and or FEV1%, nction used ent [10,15]. all the four quire lung

one another within the

id not have phone call. (with error ee whether d or over-because the cceptability an outlier.

we use the The figure

Call and the easurement

We focus mmon lung e difference mated (false ft and Top-nd to over-

stle one call.


5682

efthdfmafbncc

OfmWesaoothpim

LTgeS

CWwtoe

F(vfd

estimate the acfunction (FEV1he false negat

device, it is mofalse negative. more false negalgorithm is datfunction is bettbias towards thnoise ratio is lochannel, the falcase of voice ca

One way to qufunction is to cmeasure (FEVWe tested for eerror through asignificant effaccuracy of Spof SpiroCall mobstructed patihere are relati

perspective, itmproperly.

Lung FunctionThe bias in groundtruth luexplore the poSpiroCall.

Comparison of We used two swhistles had slio the input flo

equal variance

Figure 10. Bla(without and wversus the valufalse negativesdashes) are als

ctual value for1% < 0.8), i.e.,tives inside thore acceptable

The main reagatives for lowta driven and tter representedhe median valower for the delse negatives arall (Figure 10,

uantify the modcalculate the sta1% in this caseffects of grouna chi-square tefect of the gpiroCall (p<0.0

might degrade fuients. Althougively few falset could mean

n Estimate witthe performa

ung function ssibility of us

Two Whistle Sizes of the vorightly differentow. We perfos on the perce

and-Altman ploith whistle) for ue obtained fros are highlighteso shown.

r some patient a false negative gray boxes. to have a false

ason the systew lung functionthe population . Therefore, thue. Considerin

evices connectere slightly morTop Right).

del’s bias towaatistical effect se) on the errondtruth FEV1%

est. We found groundtruth F05). As such, tfurther if testedh the bias is e negatives, frn that patient

h Whistle ance of the of the user ping a whistle

Sizes rtex whistle in t gradient of pirmed a two-saentage error fo

ots of percent local and voice

om the clinical ed inside grey

s with low lunve. We highligIn any medic

e positive thanem currently hn is because thwith higher lune model tends ng the signal-ted over the GSre pronounced

ards higher lunof lung functio

or of the mode% on the percethat there was

FEV1% on ththe performan

d on more highonly slight an

rom a diagnostts are screen

system due prompted us for the users

our study. Boitch with respeample F-test for the four lun

error of FEV1%e call recordingspirometer. Thboxes. ±2σ (re

ng ght cal n a has he ng to

to-M in

ng on el. ent s a he

nce hly nd tic ed

to to of

oth ect for ng

functiosignificfavor betwee2.13% Considour ana

AccuraBar gpercendevice EricssoHowevtest, phighestvariancspiromFEV1%device

OutlierIn ordewhistleAltmanbetweethe spiwhistlenegativwhistlerecordiperformthe whsaw in superiolow lunby conpercenwas noaccurac

FigureSpiroC

CurvesWe noto the curves performdiagnoA tech

% gs he ed

on measures focant effect (p<of the bigger

en the two wfor PEF, FE

dering the biggalysis of Spiro

acy of Lung Fungraphs shown

ntage error of and connect

on W580i perfver, the differep>0.05). Amont for PEF, buce in PEF wa

meters. The mo%, has less thatypes.

rs and Patientser to understane results, Figurn plots of FEVen SpiroCall (wirometer measue mitigates faves inside grae comes from ings and vomance differenhistle, we elimn case of no whor screening tng function. Wnsidering the ent error througho significant effcy of SpiroCal

e 11. Two FCall with a whist

s Generated bow shift our di

shape of the serve two pu

med the effoosis by showinghnician looks a

for both whistl<0.01) of size r whistle. Th

whistles was 0EV1, FVC, anger whistle woCall only inclu

nction Measuren in Figureeach lung fun

tion type withformed the worence was not stng the lung fut it is worth

as also the higost widely usedan 8% mean e

s with Low Lungnd the directiore 10 (Bottom)V1%, displayinwith whistle) aure. From thesalse negatives. ay boxes. Mofalse positives

oice calls, thnce (F-test, p>

minate the biashistle. This metool, especially

We quantify thieffects of grouh a chi-square ffect of the groll across device

low vs. Volumetle and without

by SpiroCall iscussion from

flow-volume urposes: (1) toort sufficientlyg the descendiat the slope of

le sizes. We oon FVC and F

he percentage 0.24%, 4.22%nd FEV1%, reorked significanudes the larger

es e 9 (Right) dinction measurh a whistle. rst among all ttatistically signfunctions, the hwhile to notghest for the gd lung functionrror for three

g Function on of the bias) shows modifng percentage and the spirome plots, we shoWe highlight

ost of the errs. When comphere is no >0.05). Howe

s in the estimaeans the whistley for patients is effect of biaundtruth FEVtest. We found

oundtruth FEVes (p>0.05).

e curves gent a whistle.

m lung functioncurves. The

o evaluate if thy, and (2) toing limb of thef the FV curv

observed a FEV1%, in

difference , 2%, and

espectively. ntly better, whistle.

isplay the re for each The Sony

the phones. nificant (F-

error was te that the groundtruth n measure, of the four

present in fied Bland-

difference meter versus

ow that the t the false

ror for the aring local significant

ever, using ate that we e may be a with very

as as before 1% on the d that there

V1% on the

nerated by

n measures spirometry he patients o help in

e FV curve. e from the


5683

start of the test to PEF. This slope should be as steep as possible, indicating that the initial blast of air was truly maximal. The investigator also looks to see if the user coughs during the spirometry maneuver. Coughing makes the descending edge of the FV curve non-monotonic as the user ends up inhaling during a cough. Therefore, it is important to evaluate how SpiroCall performs in generating these curves.

Figure 11 shows example flow-volume curves generated by SpiroCall without the whistle and with the whistle; we find that the curves generated without a whistle can be unreliable. The no-whistle (green) curve in the Figure 12 (Right) has an inaccurate shape because the latter half of the effort by the patient was very quiet. When the GSM channel compressed the audio, this segment was heavily compressed and not reconstructed accurately. However, these curves can still be used for validity assessment of the efforts. The initial part of the effort is always very loud and reconstructed accurately. Therefore, the investigator can still look at the ascending slope at the start of the test. For cough information, we envision that the Hilbert envelopes of the temporal audio data can be attached along with the spirometry curves, which would make any coughs clearly visible. However, in cases where the spirometry curves are of importance, we suggest the use of a whistle. The whistle generates a direct mapping to the Flow vs. Time curve and the final Flow vs. Volume curves are usually very accurate. We recognize that a more rigorous evaluation of the spirometry curves is important. This is part of our on-going work, where we are sending all the curves generated by SpiroCall to medical practitioners for quality assessment at Spirometry 3601.

DISCUSSION SpiroCall offers two approaches to performing spirometry through a call-in service: with a vortex whistle and without. The performance of both approaches is very promising and the mean error of the four major lung function measures is 6.2%, which is well within the ATS criteria for a clinical spirometer. However, the system sometimes over-estimates lung function when used without a whistle. We believe that this limitation stems from the fact that without a whistle, the algorithm depends on the spread and variation in its training data to remove the bias in its estimation. We plan to combine the SpiroCall clinical evaluation with ongoing SpiroSmart clinical trials.

The linear relationship between flow-rate and pitch makes the vortex whistle reliable for estimating lung function measures and spirometry curves with significantly fewer false negatives and almost no bias toward high lung function. Another major advantage with the whistle is that its estimation model is generalizable across devices and channels. In fact, it calculates PEF and FEV1 directly,

1 www.spirometry360.org

without any statistical modeling. For the patients with obstructive lung impairments such as asthma and COPD, the lung function measure that changes most drastically is FEV1. If the patient only needs to track their FEV1 with fine granularity (a common practice for many patients), SpiroCall can use a much simpler computation with a whistle, without any machine learning. Moreover, it will be easier to judge a valid effort because the shape of the curve is more faithfully represented.

SpiroCall’s performance is promising as the mean performance loss due to use of the call-in service is only around 1%. The flexibility between channels and the possibility of using a whistle allows SpiroCall to make spirometry accessible. However, this only demonstrates the feasibility of sensing. It remains unclear how the user, in general, could use spirometers without any guidance from trained personnel. Although SpiroSmart tries to bridge this gap with a rich visual interface, it will be more difficult for SpiroCall to train the user. It is possible that in future work we could implement audio feedback between spirometry efforts, or have a health worker train the user before they are able to use SpiroCall independently.

CONCLUSION In order to make spirometry more accessible, it is important to remove its dependence on smartphones. We introduced SpiroCall, a combination of call-in service and a simple whistle that turns every mobile phone in the world into a spirometer. The phone sends the audio data generated during a spirometry effort over the GSM voice channel and calculates the results on a central server. Our evaluation shows that we can use SpiroCall to reliably measure lung function in low resource regions. SpiroCall’s call-in service’s mean error is comparable to a clinical spirometer and does not degrade substantially when compared to local recordings made on a smartphone. The whistle helps in improving the performance with patients with degraded lung function. SpiroCall also serves as a demonstration that researchers can perform sensing on all mobile phones, not just smartphones, by leveraging the voice channel for data transfer.

REFERENCES 1. J M Bland and D G Altman. 1986. Statistical methods

for assessing agreement between two methods of clinical measurement. Lancet 1, 8476: 307–10. Retrieved July 9, 2014 from http://www.ncbi.nlm.nih.gov/pubmed/2868172

2. Robert O Crapo, John L Hankinson, Charles Irvin, and Neil R. MacIntyre. 1994. Standardization of Spirometry. American Journal of Respiratory and Critical Care Medicine, 7.

3. Bradley Efron, Trevor Hastie, Iain Johnstone, and Robert Tibshirani. 2014. Least angle regression. The Annals of Statistics 32, 2: 407–499. Retrieved December 4, 2014 from http://projecteuclid.org/euclid.aos/1083178935


5684

4. Ericsson. 2015. Ericsson Mobility Report: On the Pulse of the Networked Society.

5. Lilian de Greef, Mayank Goel, Min Joon Seo, et al. 2014. BiliCam: Using Mobile Phones to Monitor Newborn Jaundice. Proceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous Computing 2014, ACM Press. http://doi.org/10.1145/2638728.2638803

6. Maya R Gupta, Eric K Garcia, and Erika Chin. 2008. Adaptive local linear regression with application to printer color management. IEEE transactions on image processing : a publication of the IEEE Signal Processing Society 17, 6: 936–945. http://doi.org/10.1109/TIP.2008.922429

7. R J Knudson, R C Slatin, M D Lebowitz, and B Burrows. 1976. The maximal expiratory flow-volume curve. Normal standards, variability, and effects of age. The American review of respiratory disease 113, 5: 587–600. Retrieved December 4, 2014 from http://www.ncbi.nlm.nih.gov/pubmed/1267262

8. Jiri Kroutil, Alexandr Laposa, and Miroslav Husak. 2011. Respiration monitoring during sleeping. Proceedings of the 4th International Symposium on Applied Sciences in Biomedical and Communication Technologies - ISABEL ’11, ACM Press, 1–5. http://doi.org/10.1145/2093698.2093731

9. Eric C Larson, Mayank Goel, Gaetano Boriello, Sonya Heltshe, Margaret Rosenfeld, and Shwetak N Patel. 2012. SpiroSmart : Using a Microphone to Measure Lung Function on a Mobile Phone. UbiComp’12.

10. M R Miller, J Hankinson, V Brusasco, et al. 2005. Standardisation of spirometry. The European respiratory journal 26, 2: 319–38. http://doi.org/10.1183/09031936.05.00034805

11. Dennis R. Morgan and Michael G. Zierdt. 2009. Novel signal processing techniques for Doppler radar cardiopulmonary sensing. Signal Processing 89, 1: 45–66. http://doi.org/10.1016/j.sigpro.2008.07.008

12. Rajalakshmi Nandakumar, Shyamnath Gollakota, and Nathaniel Watson. 2015. Contactless Sleep Apnea Detection on Smartphones. Proceedings of the 13th

Annual International Conference on Mobile Systems, Applications, and Services - MobiSys ’15: 45–57. http://doi.org/10.1145/2742647.2742674

13. Michael R Neuman. Vital signs: heart rate. IEEE pulse 1, 3: 51–5. http://doi.org/10.1109/MPUL.2010.939179

14. Sohail Rafiqi, Chatchai Wangwiwattana, Jasmine Kim, Ephrem Fernandez, Suku Nair, and Eric C Larson. 2015. Pupil Ware : Towards Pervasive C ognitive Load Measurement using Commodity Devices. Proc. 8th Int. Conf. PErvasive Technol. Relat. to Assist. Environ.

15. Alessandro Rubini, Andrea Parmagnani, and Michela Bondì. 2011. Daily variations in lung volume measurements in young healthy adults. Biological Rhythm Research 42, 3: 261–265. http://doi.org/10.1080/09291016.2010.505456

16. Hiroshi Sato and Kajiro Watanabe. 2000. Experimental study on the use of a vortex whistle as a flowmeter. IEEE Transactions on Instrumentation and Measurement 49, 1: 200–205. http://doi.org/10.1109/19.836334

17. Bernard Vonnegut. 1954. A Vortex Whistle. The Journal of the Acoustical Society of America 26, 1: 18–20.

18. Kajiro Watanabe and Hiroshi Sato. 1994. Vortex Whistle as a Flow Meter. Proc. Advanced Technologies in Instrumentation and Measurement, 1225–1228.

19. World Health Organization. 2015. Chronic obstructed pulmonary diseases (COPD). Retrieved September 9, 2015 from http://www.who.int/mediacentre/factsheets/fs315/en/

20. Hui Zou and Trevor Hastie. 2005. Regularization and variable selection via the elastic net. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 67, 2: 301–320. http://doi.org/10.1111/j.1467-9868.2005.00503.x

21. 佐藤浩志, 大原昌幸, 渡辺嘉二郎, and 佐藤秀昭. 1999. Application of the Vortex Whistle to the Spirometer. 計測自動制御学会論文集 35, 7: 840–845.


5685

spirocall: measuring lun g function over a phone call · ngs. the spiro mulative volu le lung funct...

Documents