1080 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 4, APRIL 2013

Cuffless Differential Blood Pressure EstimationUsing Smart Phones

Vikram Chandrasekaran, Ram Dantu∗, Srikanth Jonnada, Shanti Thiyagaraja, and Kalyan Pathapati Subbu

Abstract—Smart phones today have become increasingly pop-ular with the general public for their diverse functionalities suchas navigation, social networking, and multimedia facilities. Thesephones are equipped with high-end processors, high-resolutioncameras, and built-in sensors such as accelerometer, orientation-sensor, and light-sensor. According to comScore survey, 26.2% ofU.S. adults use smart phones in their daily lives. Motivated bythis statistic and the diverse capability of smart phones, we fo-cus on utilizing them for biomedical applications. We present anew application of the smart phone with its built-in camera andmicrophone replacing the traditional stethoscope and cuff-basedmeasurement technique, to quantify vital signs such as heart rateand blood pressure. We propose two differential blood pressureestimating techniques using the heartbeat and pulse data. The firstmethod uses two smart phones whereas the second method replacesone of the phones with a customized external microphone. We es-timate the systolic and diastolic pressure in the two techniquesby computing the pulse pressure and the stroke volume from thedata recorded. By comparing the estimated blood pressure valueswith those measured using a commercial blood pressure meter, weobtained encouraging results of 95–100% accuracy.

Index Terms—Blood pressure, cuffless, finger pulse, mobile cam-era, mobile phone, vascular transit time.

I. INTRODUCTION

M EASUREMENT of vital signs in a human is an arduoustask when sudden dizziness or fainting occurs during

unexpected situations. The only information that we are awareof is that these occurrences are the most common symptoms oflow blood pressure. Blood pressure, the amount of force appliedon the walls of the arteries when the blood is forced throughoutthe body, depends on factors such as the amount of blood inthe body, the pumping rate of the heart, the flexibility of thearterial walls, and the resistance to blood flow due to the size of

Manuscript received September 7, 2011; revised December 2, 2011,January 29, 2012, March 15, 2012, and July 19, 2012; accepted July 20, 2012.Date of publication August 1, 2012; date of current version March 15, 2013.This work was supported in part by the University of North Texas InstitutionalReview Board (IRB) and in part by the National Science Foundation underGrant CNS-0751205 and Grant CNS-0821736. Asterisk indicates correspond-ing author.

V. Chandrasekaran, S. Jonnada, S. Thiyagaraja, and K. P. Subbu arewith the Network Security Laboratory, Department of Computer Sci-ence, University of North Texas, Denton, TX 76203 USA (e-mail:vikram.chandrasekaran@unt.edu; srikanthjonnada@my.unt.edu; srt0106@unt.edu; kp0124@unt.edu).

∗R. Dantu is with the Department of Computer Science and Engineering,University of North Texas, Denton, TX 76203 USA (e-mail: rdantu@unt.edu).

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TBME.2012.2211078

the arteries. The blood pressure of a human varies continuouslydue to physical activity, medication, anxiety, and emotions. Thebody has unique mechanisms to regulate a person’s blood flow;whenever a person’s blood pressure drops, the heart rate in-creases to pump more blood and the arterial walls contract toincrease the blood pressure.

Blood pressure is typically measured using a mercury sphyg-momanometer. However, to check a person’s blood pressureduring unexpected situations, there is a need for a portable, con-venient device or apparatus. Despite the availability of digitalwrist and arm blood-pressure meters, most people do not carrythese devices during their daily commute to the work place,gymnasiums, recreational facilities, etc. The only device whichis omnipresent with people in all walks of their life is the mo-bile or smart phone. The mass production of smart phones withbuilt-in sensors such as an accelerometer, a camera with flash,and a microphone sensor has motivated us to tap their benefits tothe biomedical domain. Although Scully et al. [1] developed amethodology for estimating the vital signs using a smart phoneit does not include the blood pressure estimation.

We hereby propose a mobile phone-based blood pressure es-timation technique. In other words, we replace the traditionalcuff and stethoscope traditionally used for blood pressure mea-surement, by recording the heart sounds and finger pulse in twoways: 1) employing the microphone and camera built-in twosmart phones, and 2) a customized external microphone and asingle smart phone’s camera. Finger pulse is recorded by imple-menting photoplethysmography technology in smart phones asdescribed by Jonathan et al. [2].

The estimation technique consists of three phases. The firstphase involves locating an appropriate spot on the chest forrecording heart sounds. This audio data are used to calculatethe heart rate. The second phase consists of two methods forrecording. In the first method, two clock synchronized mobilesare used, one for recording the heart sounds and other for record-ing finger pulse. In the second method, we use a single mobilewith external microphone for recording both heart sounds andfinger pulse. The third phase, i.e., the blood pressure estimationphase involves calculation of systolic pressure, pulse pressure,and the diastolic pressure using the data measured.

II. METHODOLOGY

The process of recording heart sounds and finger pulse usingsingle mobile and two mobiles are discussed in detail in thissection. The localization of heart beat and pulse, pulse detection,and heart rate calculation for both single mobile and two mobilemethods are discussed in the next sections.

0018-9294/$31.00 © 2012 IEEE

CHANDRASEKARAN et al.: CUFFLESS DIFFERENTIAL BLOOD PRESSURE ESTIMATION USING SMART PHONES 1081

Fig. 1. Blood pressure estimation test setup: Pulse signals are obtained usingthe video recording application when the finger is placed over the lens and heartsounds are recorded simultaneously using the audio application by placing theother phone over the chest as shown.

A. Setup of Two Mobiles for Blood Pressure Estimation

In the first technique, the setup starts with synchronization oftime between two mobile phones using Bluetooth. After syn-chronization, the camera on one mobile records the pulse inthe finger at 24 frames/s (fps) via a video recording applica-tion. An LED flash lights up when the application starts andthen the start-time of video recording is saved in the phone’sSD card. Similarly, the second mobile starts recording the heartsounds via the audio recording application with the phone’smicrophone held close to the chest. Care has to be taken thatthe microphone’s opening is held tightly to the skin over thechest to avoid recording external noise. Fig. 1 shows the overallprocedure of the differential estimation technique.

For synchronizing the clocks on both the mobile phones, wedesigned a Bluetooth-based synchronization protocol. This pro-tocol is similar to the precision time protocol-IEEE 1588 [3]used in wired networks. The synchronization procedure followsa master–slave architecture. The mobile phone which receivesthe synchronization messages acts as the master and the onesending acts as the slave. We employed the Nexus One and HTCHero mobile phones running Android operating system. TheNexus One was equipped with a Broadcom BCM4329 chipsetsupporting Bluetooth 2.1+EDR (extended data rate technol-ogy) along with 802.11n WiFi and the Hero supports Bluetooth2.1+EDR along with a 802.11b/g WiFi connectivity. The phonescommunicated via the Android BlueZ [4], a Bluetooth 2.1 com-patible stack capable of running on any Bluetooth chipset. TheNexus One used for video recording, acted as the master andthe Hero, used for heart beat recording, acted as the slave. Thesynchronization procedure consists of two steps as explainedbelow.

1) Slave device synchronization to master clock: The mas-ter device’s timestamp is obtained through Bluetooth andis fixed on the slave device. This simple method syn-chronizes time on two devices, but differential estimationof blood pressure requires time calculations precise tomilliseconds. Ignoring transmission delays will lead todegradation in the system’s accuracy. Hence, transmis-sion delays have to be calculated during the procedure.

Fig. 2. Flow of events in Synchronization: TS + i is the sending time in slavedevice. T ′

S + i is receiving time in slave device. TR+ i is the receiving timein master device. T ′

R+ i is sending time in master device. N is the maximumnumber of cycles the process can continue

This requires the timestamps of both the systems. Fig. 2depicts the working of the synchronization process. Thefirst step is the standard establishment of a Bluetooth con-nection after pairing with the device. Once the connectionis established, the slave device sends a timestamp to themaster and stores it as TS . As soon as the master devicereceives the timestamp, it sends back its own timestampto the slave device. This returned timestamp is recordedas TR along with the timestamp when TR was received bythe slave device as T ′

S . The process is repeated for 30 sand the timestamps are recorded. This process is followedby the client clock synchronization module in the slavedevice. This module has four major operations: estimat-ing the roundtrip time, estimating the offset, calculatingthe master time, and setting the slave clock. The roundtriptime is given by

RTT = T ′S − TS . (1)

The one-way transit time between the terminals is one-halfof the roundtrip time given by

Tt =RTT

2. (2)

The offset between the two devices Ot is given by differ-ence in timestamp of the master and slave device

Ot = TS − (TR − Tt). (3)

Subtracting the one-way transit time from the receiv-ing time in the master device will produce the accuratetime difference between the slave and master device. The

1082 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 4, APRIL 2013

Fig. 3. Design of customized external microphone: when the diaphragm isplaced on subject’s chest, body sounds vibrate the diaphragm creating acousticpressure waves which travel through the hollow tube and the microphone, whichis fixed at the other end of hollow tube, captures the acoustic waves and transfersthem to the mobile phone.

master clock time is estimated by subtracting the offsetwith the slave clock.

In order to overcome the overhead of measuring at-a-timepulse rate and heart beat of an individual to estimate bloodpressure, the existing system is improvised to a single mobileapproach.

B. Setup of Single Mobile Phone for Blood Pressure Estimation

In this setup, we record heart sounds with a customized exter-nal microphone and finger pulse at 24 fps via a video and audiorecording application. An LED flash lights up when the applica-tion starts and then the start-time of video and audio recordingare saved in the phone’s SD card.

The design of the external microphone consists of an acousticstethoscope diaphragm, hollow tube, 3.5 mm mini-plug con-denser microphone with frequency response 50–18,000 Hz anda 3.5-mm microphone adapter for mobile phones. We fix thediaphragm to one end of hollow tube and microphone to otherend of the hollow tube. Fig. 3 shows the customization of exter-nal microphone. This customized microphone replaces the needfor the second mobile which is used to measure heart sounds inFig. 1.

III. INTERWORKING OF COMPONENTS FOR BLOOD

PRESSURE ESTIMATION

Before the measurement starts, we have to identify the bestlocation on chest for heart beats to attain good quality of read-ings. The localization of heart sounds, analysis of finger pulse,and calculation of heart rate are explained in Sections III-Aand III-B. The process of correlating the heart beats and fingerpulse for analyzing the blood pressure is clearly explained inSection III-C.

A. Localization of Heart Beat

Localizing a heart beat is a challenging task, but necessaryfor accurately determining blood pressure. A brief descriptionof heart sounds is as follows. Heart sounds are produced withthe opening and closure of heart valves. The heart producesmainly four types of sounds [5] in one heart cycle denoted asS1, S2, S3, and S4. The first heart sound (S1, lub) is producedby the atrioventricular valves (i.e., mitral and tricuspid), thesecond heart sound (S2, dub) is produced by the semilunarvalves (i.e., aortic and pulmonary valve). The third and fourth

Fig. 4. Heart sound locations: S1 (lub) is produced by mitral (M) and tricuspid(T) values; S2 (dub) is produced by aortic (A) and pulmonary (P) values

TABLE ILOCATION OF VALVE ON CHEST

Pulmonary valve second intercostal space left upper sternal border

Aortic valve second intercostal space righ upper sternal border

Mitral valve fifth intercostal space medial to left midclavicular line

Tricuspid valve fourth intercostal space lower left sternal border

heart sounds are produced only in some rare conditions due togallop. In this paper, we are interested in recording only the firstand second heart sounds. Experiments were conducted on thechest region to find the best spot for obtaining recordable heartsounds. As indicated in Fig. 4, the four locations of the valveswere identified and selected for recording purposes. Based onthe experiments, the heart sounds captured from the microphonewere heard the loudest in the pulmonary region containing thepulmonary valve. Since the pulmonary valve is associated withthe S2 sound, in most cases the decibel level of S2 is higher thanthat of the S1. Table I as listed in [6] gives the location of valvesin the chest with respect to the intercostal space and sternum.

B. Pulse Detection and Heart Rate Calculation

The procedure for detecting a pulse works on the principle [7]that, every heart beat pertains to a rush of blood in the bloodvessels, even in the capillaries at the finger-tips. By placingthe finger over the camera and turning on the flash through avideo recording application, the following observations weremade. During a systolic pulse when the capillaries are rich inblood, more light was absorbed by the blood, leading to a low-reflective index and darker frame intensities. Likewise, during adiastolic pulse, most of the light was reflected leading to brightframes. The change in intensity of light passed through thefinger creating an alternative pattern of waves. These changesin intensity with time were used to obtain the pulse of a person.

The pixel information from the video was split into individualred, blue, and green components. In most of the frames observed,the prominent color was red. We split every frame of the videointo four quadrants and considered only the first quadrant for

CHANDRASEKARAN et al.: CUFFLESS DIFFERENTIAL BLOOD PRESSURE ESTIMATION USING SMART PHONES 1083

0 5 10 15 20 25 30 35 40190

195

200

205

210

215

220

225

230

235

Time (s)

Red

inte

nsity

(0−

255)

filtered intensity

(a)

0 1 2 3 4 5 6 7 8 9 10206208210212214216218220222224

Time(s)

Inte

nsity

10 11 12 13 14 15 16 17 18 19 20200

205

210

215

220

225

Time(s)

Inte

nsity

20 21 22 23 24 25 26 27 28 29 30190195200205210215220225230

Inte

nsity

Time(s)30 31 32 33 34 35 36 37 38 39 40

195

200

205

210

215

220

225

230

Time(s)

Inte

nsity

(b)

Fig. 5. (a) Total window of data. (b) Data from first quadrant split into smallertime frames.

analysis because the fluctuations were more predominant in thatregion. Then we split the frames on a fixed length window (Wt)to check for the number of peaks (n) occurring at equal intervalsof time as shown in Fig. 5. We then calculate the Heart rate (HR)given by

HR = (n ∗ 60)Wt. (4)

C. Integration of Components

This section explains how the recorded heart beats and fin-ger pulse are processed and used to estimate the systolic anddiastolic pressures. Fig. 6 illustrates the sequential execution ofevents in the estimation process. The estimation process startsby first synchronizing the system clocks of both the mobilephones. After this, each mobile starts recording the appropriatedata. The Nexus One recorded video at a rate of 24 fps and theHTC Hero recorded audio at a sampling rate of 8 kHz. Boththe mobile phones must start recording within 10 s after clocksynchronization. Letting the application to run for a long time,creates a drift in the system clock as the processor is responsiblefor maintaining the system time. Moreover, when the CPU isloaded with high-priority processes, the processor cannot main-tain the clock’s accuracy.

Fig. 6. Working of cuffless differential estimation of blood pressure.

Algorithm 1 Calibration for Synchronous start of audio andvideo dataINPUT:Timestamps of Audio and Video Start (SA and SV )OUTPUT:Synchronous audio dataSyncStart = SA − SV

if SyncStart ≥ 0 thenAdd Null data for SyncStart milliseconds towards start ofthe audio data set

end ifif SyncStart < 0 then

Delete data for SyncStart milliseconds from the start ofthe audio data set

end if

In our study, all the measurements were taken for a duration of20 s with the video starting roughly 1 s after the audio. The datacollected were processed offline in MATLAB. The first step wasthe denoising of data and resampling. Audio data collected fromthe mobile phone were passed through a 15th-order band-passfilter which allows frequencies only between 10–250 Hz sincemost heart sounds recorded are present within the 10–250 Hzfrequency band. For computational effectiveness, the filteredaudio was resampled to a lower rate of 1 kHz. Similarly, thevideo frames were processed as explained in Section III-B togive the value of red intensity.

The next step was determination of peaks from the data andidentifying appropriate peaks in the video corresponding to theaudio signal. To identify peaks, the waveforms had to be alignedsince audio and video data were not synchronized with start time.To achieve this, the starting times of the recording were storedon the SD card of the two phones, respectively, as time stampsfor calibration. For instance, let us suppose the starting time ofthe audio data was SA and the starting time of the video data wasSV . The difference in the time stamps will give the amount oftime one data was ahead/lagging the other. Algorithm 1 explainsthe procedure followed in calibrating the two devices. Since thesampling frequency for audio was 1 kHz, adding and deletingdata points was easier and more precise in audio than video at24 fps. Fig. 7 Illustrates the aligned data.

After performing the alignment, a peak detection algorithm[8] was applied on both the waveforms. The algorithm uses apeak function for each sample in the waveforms defined as themean of the maximum values between the left and right set

1084 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 4, APRIL 2013

Fig. 7. Identification of Peaks: sys represents the systolic peak of the pulse.S1 and S2 are the first and second heart sounds, respectively. VTT is identifiedby time difference between the dotted lines.

of neighboring points for each sample. Further, based on thecalculated mean and standard deviation for each of the peakfunction values, the local peaks were filtered. These local peaksare formed smaller in magnitude next to the systolic peaks asseen in Fig. 7. This resulted in identifying one significant peakwithin a window revealing a clear existence of a heart beat.

IV. MATHEMATICAL ANALYSIS

This section describes the method of calculating systolic anddiastolic pressure values using the preprocessed heart beats andfinger pulse.

A. Systolic Pressure

The times of arrival of S1 and the corresponding systolic peakare noted. The difference in these two time of arrivals yieldeda parameter called vascular transit time (VTT). VTT is definedas the time taken by the blood to travel from the heart to anextremity of the body for one stroke of the heart. Foo et al. [9]state that the change in systolic pressure can be derived fromthe change in VTT with respect to a reference value as shownin (5). Hence, the reason for calculating VTT

�Ps = −0.425 ×�VTT. (5)

Based on (5), the systolic pressure values corresponding theVTT can be generated. Thus, the estimated systolic pressureusing two mobiles is given by

Ps = −0.425 × VTT + 214. (6)

To derive the relation between systolic pressure and VTT weperformed linear curve fitting between the estimated VTT andmeasured systolic pressure values to obtain a regression relation.The above constant 214 is obtained from the regression relation.We have validated this constant for more than 500 readings.

Fig. 8. Heart beat data recorded with the microphone: the two points markedin the waveform correspond to S1 and S2 sounds. The time difference betweenthese two sounds gives ejection time.

B. Pulse Pressure and Stroke Volume

The pulse pressure Pp and the stroke volume SV were com-puted as follows. For stroke volume, we used (7) formulatedin [10] as

SV(mL) = −6.6 + 0.25 × (ET − 35)

− 0.62 × HR + 40.4 × BSA − 0.51 × Age (7)

where ET(ms) is the ejection time and BSA is the body surfacearea given by

BSA = 0.007184 × Weight0.425 × Height0.725 .

Ejection time is defined as the time of ejection of blood fromthe left ventricle beginning with the opening of the aortic valveand ending with closing of the aortic valve. The ejection timewas estimated graphically from the heart sound recorded asshown in Fig. 8. The identification of the two sounds namely S1accompanied by opening of the aortic value and S2 accompaniedby closing of the aortic value was explained in Section III-A.

As shown in Fig. 8, the time difference between S1 and S2sounds recorded using the microphone is taken as the ET.

With the computed stroke volume, the pulse pressure wascalculated in units of mmHg using (8) formulated in [11]

Pp =SV

(0.013 × Wt − 0.007 × age − 0.004 × HR) + 1.307.

(8)

C. Diastolic Pressure

Having obtained Ps and Pp , the diastolic pressure Pd wascalculated from the following equations [11]:

Ps = Pm +23Pp (9)

Pd = Pm − Pp

3. (10)

Subtracting (10) from (9) and rearranging, we get

Pd = Ps − Pp. (11)

CHANDRASEKARAN et al.: CUFFLESS DIFFERENTIAL BLOOD PRESSURE ESTIMATION USING SMART PHONES 1085

(a) (b)

Fig. 9. Sharp and flat peaks: (a) sharp peak occurs only at a single time stampwhile (b) flat peak occurs over a continuous time stamp.

Hence, using the preprocessed heart beat and finger pulsesignals systolic and diastolic pressure values were calculatedusing the equations described in this section.

V. RESULTS

In this section, the accuracy of differential estimation usingsingle mobile is discussed in detail followed by sensitivity anal-ysis, impact of age, height, and weight on systolic pressure.Further the performance analysis of synchronization protocoland microphone is discussed.

A. Accuracy of Differential Estimation Using Single Mobile

The accuracy of the results relies entirely on the effectiveestimation of VTT from the graphs. Identification of the timeinstance of S1 sound from the audio was an easy task with properfiltering of the signal, but the selection of the correspondingpeaks in the video intensity was difficult in some situations dueto presence of flat peaks instead of a sharp for a systolic pulsein the video intensity plot. The flat peak pattern appears due toinappropriate pressure applied by the finger and misplacementof finger while placing on the camera. Fig. 9(a) shows a plotwith sharp peaks and Fig. 9(b) shows a pattern of flat peaks. Asample data set was considered and the VTT was obtained usingboth sharp and flat peaked data points. Table II shows the valuesof VTT and Ps computed from data consisting of 1) ten sharppeaks (SP); 2) five sharp peaks and five flat peaks (SP + FP);and 3) ten flat peaks (FP). For verifying the Ps estimated usingthis method, we measured the blood pressure as 125 mmHg fora single subject using a commercial meter. From Table II, it wasdetermined that the presence of sharp peaks in the data resultedin higher accuracy when compared to flat peaks. Hence, for anaccurate measurement of blood pressure, the sharp peaks playan important role.

The diastolic pressure is approximated based on the amountof blood in the person for his age, calculated from ET. Theaccuracy may change in the case of blood loss or deficiency ofblood due to abnormality. The VTT was analyzed based on theaudio and video samples measured together. For every singlepulse in the data set, a VTT can be derived. However, the VTTcan change on every pulse since there is a continuous variation ofpressure in a human throughout the day. By using the proposedmethod, the systolic pressure was determined with a single VTTvalue but the probability of erroneous results was high. Thus,

TABLE IISHARP VERSUS FLAT PEAK: WITH AN ACTUAL Ps OF 125, THE ESTIMATED Ps

(MMHG) USING ONLY SHARP PEAKS RESULTED IN HIGHER ACCURACY AS

COMPARED WITH SIGNALS INCLUDING FLAT PEAKS

In column 4, (F) indicates flat peak.

SP SP + FP FPVTT Ps VTT Ps VTT Ps

215 122 220 (F) 121 230 116207 126 224 (F) 119 220 121202 128 213 123 226 118212 124 216 122 235 114215 122 226 (F) 118 226 118214 123 215 122 221 122215 122 215 122 225 118212 124 213 123 223 119215 122 233 (F) 115 224 119207 126 222 (F) 120 234 115

Average 211 124 220 120 226 118Accuracy (%) 99 96 94

25 30 35 40 45 50 55 60

40

60

80

100

120

140

Time (s)

Sys

tolic

Pre

ssur

e (m

mH

g)

Maximum Pressure 126 mmHg

Minimum Pressure 102 mmHg

Fig. 10. Variability of Ps recorded using the proposed method for 1 min.: thesample data set varied within a range of 24 mmHg. The median of the pressurevalue resulted in 113 mmHg which was exactly the same from the blood pressuremeter for that instant.

TABLE IIIVARIATION OF ESTIMATED (Es ) SYSTOLIC PRESSURE (Ps ) FROM MEASURED

(Md ) BASED ON SELECTION OF NUMBER OF DATA POINTS

FROM A DATA SET, UNITS IN mmHg

Sample Md Ps 1 point 5 points 10 points1 110 105 112 1102 115 110 110 1123 108 101 112 1074 104 115 114 111

for all the analyzed data sets, the median of the VTT was takento compute the blood pressure. Fig. 10 shows a plot for systolicpressure values for a duration of 60 s in a single data set.

The results from Table III, introduce the problem involvedin the selection of number of data points. Although the systolicpressure can be estimated by selecting a single data point, theaccuracy of the values will be very low. Table III contains fourdata sets showing the pressure values with different number ofdata points selected. A single point was taken at random froma data set and compared against the actual pressure value. In asimilar manner, the pressure value was computed by selectingfive and ten points. All the plots showed results close to actualpressure, but the closeness increased with increase in the numberof data points selected.

From the Table III, we can infer that the selection of at leastten data points from the dataset yielded closer values of pressure.

1086 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 4, APRIL 2013

100 105 110 115 120 125 130 135 140−8

−6

−4

−2

0

2

4

6

Average blood pressure from two methods (mmHg)

Diff

eren

ce in

blo

od p

ress

ure

mea

sure

d us

ing

trad

ition

al a

ndsm

art p

hone

(m

mH

g)

datameanmean±2SD

(a)

55 60 65 70 75 80 85 90 95 100−6

−4

−2

0

2

4

6

8

Average blood pressure from two methods (mmHg)

Diff

eren

ce in

blo

od p

ress

ure

mea

sure

d us

ing

trad

ition

al a

ndsm

art p

hone

(m

mH

g)

datameanmean±2SD

(b)

Fig. 11. Bland–Altman Plot: (a) systolic pressure and (b) diastolic pressure.The blood pressure estimated using single smart phone lies within mean 2SDindicating the similarity to the traditional method.

A Bland–Altman plot for single mobile method was con-structed with 50 readings. The plot is shown in Fig. 11. Ourproposed method using single mobile yielded an accuracy of95% confidence intervals of ±2.4473 mmHg in systolic pres-sure and 95% confidence intervals of ±1.7073 mmHg in dias-tolic pressure from a data set comprising of 500 measurementsobtained from both male and female between an age range of21–55 years.

1) Accuracy Levels Due to Effect of Noise and Motion Arti-facts: We employed five subjects and each subject performedthree experiments separately.

a) Still with single mobile: The subjects placed their fin-ger over the camera lens and LED flash while the diaphragmof customized external microphone is placed over their chest.The video–audio recording application records the finger pulseat a frame rate of 24 fps, 720 × 480 resolution and audio at asampling rate of 8 kHz for 20 s. The subjects were instructednot to make any movements during the recording of audio andvideo.

b) Movements with single mobile: The same subjectswere used to repeat the experiment. The setup of the experi-ment is same as the previous experiment. However, during therecording, the subjects were allowed to make small randomhead and chest movements such as tilting the head and movingthe chest back and forth. Such movements cause the noise and

motion artifacts to be induced into the recording of audio andvideo. There was no fixed time for the movements and hencewere not tabulated.

c) Traditional meter: The same subjects’ blood pressurewas measured using the sphygmomanometer.

The measured and estimated blood pressures are tabulated inTable IV. From the table, we can see that although the move-ments caused variation in blood pressure, the accuracy still re-mained closed to that obtained while at a still position.

B. Sensitivity Analysis for Cuffless Estimation Methodin Mobiles

In this section, we discuss whether factors such as age, height,and weight affect the system’s sensitivity for the systolic anddiastolic estimation.

1) Systolic Pressure: The tests were performed with subjectsof varying age, height, and weight as listed in Table V.

By obtaining the VTT and Ps for the five subjects as explainedin Section III, we compared the estimated values of Ps from VTTwith the pressure measured using a meter. Fig. 12 shows a plotof systolic pressure versus VTT. The meter-based readings arejoined by a straight dark line in the center. The points labeledas “Data” refer to different individuals. The results obtained byapplying (6) to all the individuals regardless of the age, height, orweight for Ps computation, yielded accuracies of above 95%.Based on these results, we can infer that factors such as age,height, and weight did not impact the end result of proposedsystem. We also observed that, for an increase of 10 ms of VTT,the systolic pressure dropped linearly by 4.25 mmHg.

The method has not been tested with people having a historyof cardiac disorders. The results and accuracy could also varyfor people with ailments and disabilities due to changes in theirvascular system.

2) Diastolic Pressure: As the systolic pressure values areobtained only from VTT, the sensitivity analysis was done ondiastolic pressure by studying the relation between the diastolicpressure and a person’s height and weight. Fig. 13 shows a slightincrease in the diastolic pressure of a person with increase ofheight and weight. From the plot, it can be seen that there is nomajor variation in the diastolic pressure for different values ofheight and weight.

C. Performance Analysis

The accuracy of results depends on the performance of a sys-tem under any conditions. This section shows the performanceanalysis of synchronization protocol and microphone.

1) Bluetooth synchronization: As explained in Section II-Athe two mobiles are synchronized through Bluetooth. Synchro-nization in time is essential to obtain appropriate S1 peak forthe corresponding video of systolic pulse data.

To estimate the performance of the Bluetooth synchronizationprotocol, we assumed that the one-way network delay is thesame on both sides, i.e., the time taken for a message from amaster to a slave is the same as that from a slave to the master.A small test for the effectiveness of the synchronization processwas done. Because it was hard to have a centralized time for

CHANDRASEKARAN et al.: CUFFLESS DIFFERENTIAL BLOOD PRESSURE ESTIMATION USING SMART PHONES 1087

TABLE IVEFFECT OF NOISE AND MOTION ARTIFACTS ON BLOOD PRESSURE, UNITS IN MMHG

Traditional Single mobileSubject meter Movements Still

Ps Pd Ps Accuracy (%) Pd Accuracy (%) Ps Accuracy (%) Pd Accuracy (%)1 118 70 114 96.61 69 98.57 116 98.31 69 98.572 116 60 114 98.28 60 100 117 99.15 61 98.363 130 85 133 97.74 87 97.70 128 98.46 84 98.824 121 68 124 97.58 70 97.14 121 100 71 95.775 98 57 101 97.03 54 94.74 101 97.03 58 98.28

TABLE VSUBJECT INFORMATION

Subject Age Height(ft) Weight(lbs)Subject 1 22 6.0 149Subject 2 28 6.1 142Subject 3 23 5.8 140Subject 4 25 5.8 145Subject 5 24 5.6 125

215 220 225 230 235 240 245 250 255 260 265

105

110

115

120

125

VTT (ms)

Sys

tolic

Pre

ssur

e (m

mH

g)

5 mmHg

5 mmHg

Subject 1

Subject 5

Subject 3

Subject 2

Subject 4

Region between representsdata with ~95% accuracyActual pressureEstimated pressure

Fig. 12. Shows different subjects follow the same data pattern given by(6). Thus, the proposed method is not affected by age, height, and weight ofindividual.

verification of synchronization process of two mobile phones,performance was measured using alternative method as follows.

1) The devices to be synchronized for the estimation werelabeled as Device 1 and Device 2.

2) Device 1 was made to act as the master device and syn-chronized making note of the offset.

3) Device 2 was made to act as the slave device and synchro-nized making note of the offset.

4) When the offset value is 0, the clocks between the devicesare perfectly synchronized.

5) Steps 2 and 3 could be repeated to verify zero offset valuein both directions.

The performance test was conducted on two pairs of devices:Nexus One versus Nexus One and Nexus One versus HTC Hero.Hero and Nexus One do not have a common hardware configu-ration. Hence, there was a tradeoff in the network delay. Basedon this experiment, we found that the lower the offset time be-tween devices, the greater the effectiveness. The offset variedbetween ±1 ms, which showed an effective synchronization ofthe clock times in both devices.

Fig. 13. Change in diastolic pressure with changes in height and weight.

A Bluetooth device can transmit data at a rate of 780 kb/saccording to the IEEE 802.11 standards. The effective transmis-sion and reduction of the RTT relies on the size of messagesbeing sent and the processor capability to handle multiple re-quests, interference, and attenuation of signal. In this paper, thedevices were placed as close as possible to eliminate the delayintroduced due to signal issues. Also the message size beingtransmitted was very small for a bandwidth of 780 kb/s. There-fore data loss due to retransmissions or transmission delays dueto buffer overflows did not occur in the system.

2) Microphone Performance: The detection of heart soundsrelies on the microphone’s ability to pick up low-frequency, low-amplitude audio signals. Every mobile phone in the market hasdifferent hardware specifications. Two brands of phones maynot have the same kind of hardware. However, a device with thebest microphone sensitivity could help in recording the heartbeats accurately. In this section, we provide a comparison ofmicrophones on three devices.

First, we performed a test to determine the frequency responseof the microphone on Nexus One, HTC Hero, and HTC Leg-end. The test occurred in a closed environment with ambientnoise and few audio activities. The three phones recorded thesame audio simultaneously. The recorded audio was processedin MATLAB to derive a frequency response.

Fig. 14 shows a flatter response curve for Hero than Leg-end and Nexus One. Legend has the highest power in the low-frequency spectrum followed by Hero and Nexus One. Also, theoccurrence of peaks in the audio data due to external disturbancewas seen only in Hero and Legend, where as it was eliminated inNexus One by the hardware noise cancellation available on thephone. Thus, heart beats were more accurately detected usingLegend or Hero than Nexus One.

1088 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 60, NO. 4, APRIL 2013

0 500 1000 1500 2000 2500 3000 3500−30

−20

−10

0

10

20

30

40

50

60

Freq (Hz)

Pow

er (

dB)

HTC LegendHTC HeroNexus One

Fig. 14. Microphone frequency response for three different mobile phones(Nexus One, HTC Hero, HTC Legend).

VI. LIMITATIONS AND FUTURE WORK

A. Limitations

Some technical and social issues arise which need to be ad-dressed when considering the use of cuffless differential estima-tion of blood pressure.

1) The method of measuring pulses with the video intensity isnative without much filtering. Hence, appearance of sharppeaks was hard to detect with existing cameras. Avail-ability of high-quality built-in cameras may effectivelyincrease the accuracy and reliability of the device.

2) The hardware specifications for all devices are not stan-dard. Device incompatibilities will most likely occur. Forexample, the LED flash available in phone may not bepresent in all devices as standard. Also, software versiondifferences make the applications to work only on certaindevices.

3) The two mobile differential estimation test requires theuse of rooted phones which are popular only among de-velopers. API changes have to be made, to incorporatechanges in these phone’s system or the application shouldbe developed as a core system process to be used by allpublic users.

4) Heart beat recording in two mobile differential estimationrequires placement of the microphone over the bare chest.Taking measurements with the shirts on does not givedesirable results. Hence, the measurement may not bepossible in public.

5) Our methodology has not been tested on people with Car-diac disorders.

6) The accuracy of estimation technique was low when thetests were performed with the subject lying flat on theground. Currently, the subject needs to be sitting up rightwith the arm holding the camera in level to the heart.

B. Future Work

Future work would include implementing the pulse detectionin real time with the ability of the mobile phone to identify in-correct holding of the phone and instruct a proper way. We also

plan to implement a real-time filtering of audio heart sounds ca-pable of detecting beats. The accuracy results were obtained byusing an off-the-shelf blood pressure meter. However, the datafrom them are not accurate either. Hence, we plan on using morereliable source of measurement for our accuracy calculations.

VII. CONCLUSION

This paper describes a cuffless differential blood pressure es-timation technique using two mobiles. Further as an extension,we also described a single mobile-based technique showing thefeasibility of measuring blood pressure using only one smartphone. We then discussed the mathematical formulations usedto estimate the systolic and diastolic pressure. By obtaining ac-curacies between 95–100%, we showed with this technique thatubiquitous mobile phones could be used by people to measureblood pressure. This study could very well be used for normaland nonpathological cases.

REFERENCES

[1] C. Scully, J. Lee, J. Meyer, A. Gorbach, D. Granquist, Fraser, Y. Mendel-son, and K. Chon, “Physiological parameter monitoring from opticalrecordings with a mobile phone,” IEEE Trans. Biomed. Eng., vol. 99,no. 5, pp. 749–898, 2011.

[2] J. E and L. Martin, “Investigating a smartphone imaging unit for photo-plethysmography,” Physiol. Meas., vol. 31, no. 11, pp. N79–N83, 2010.

[3] S. Lv, Y. Lu, and Y. Ji, “An enhanced IEEE 1588 time synchronization forasymmetric communication link in packet transport network,” Commun.Lett., vol. 14, no. 8, pp. 764–766, 2010.

[4] Bluez: Bluetooth Stack for Android. [Online]. Available:http://source.android.com/porting/bluetooth.html.

[5] L. S. Bickley, Bate’s Guide to Physical Examination and History Taking,8th ed. Baltimore, MD: Lippincott Williams & Wilkins, 2002.

[6] Heart Sounds Types and Locations. [Online]. Available:http://en.wikipedia.org/wiki/Heart_sounds.

[7] K. Banitsas, P. Pelegris, T. Orbach, D. Cavouras, K. Sidiropoulos, andS. Kostopoulos, “A simple algorithm to monitor hr for real time treatmentapplications,” in Proc. Int. Conf. Inf. Technol. Appl. Biomed., Nov. 2009,pp. 1–5.

[8] G. K. Palshikar, “Simple algorithms for peak detection in time-series,” inProc. 1st Int. Conf. Adv. Data Anal., Business Anal. Intell., 2009.

[9] J. Y. A. Foo, C. S. Lim, and P. Wang. (2006). “Evaluation of blood pressurechanges using vascular transit time,” Physiol. Meas. [Online]. 27(8), p.685, Available: http://stacks.iop.org/0967-3334/27/i=8/a=003.

[10] J. N. Cohn, S. Finkelstein, G. McVeigh, D. Morgan, L.LeMay, J. Robinson, and J. Mock. (1995). “Noninvasivepulse wave analysis for the early detection of vascular dis-ease,” Hypertension. [Online]. 26(3), pp. 503–508, Available:http://hyper.ahajournals.org/cgi/content/abstract/26/3/503.

[11] J. Alfie, G. D. Waisman, C. R. Galarza, and M. I. Camera. (1999).“Contribution of stroke volume to the change in pulse pressure pat-tern with age,” Hypertension. [Online]. 34(4), pp. 808–812, Available:http://hyper.ahajournals.org/cgi/content/abstract/34/4/808.

Vikram Chandrasekaran received the Bachelor’sdegree from Anna University, Chennai, India, in 2008and the Master’s degree in computer science fromUniversity of North Texas, Denton, Texas, in 2010.

He is the chief system architect for Likeminds,Inc., in New York. His research interests include ap-plication and information security, wireless and com-puter networking, VoIP, cryptography, social networkanalysis, mobile applications for health care.

CHANDRASEKARAN et al.: CUFFLESS DIFFERENTIAL BLOOD PRESSURE ESTIMATION USING SMART PHONES 1089

Ram Dantu received the B.Eng. degree from theMadras Institute of Technology, Chennai, TamilNadu, India, the M.Eng. degree from the Madras Uni-versity, Chennai, and the Ph.D. degree from Concor-dia University, Montreal, QC, Canada.

Since his Ph.D., he has more than 20 years of expe-rience in the networking industry, where he workedfor Cisco, Nortel, Alcatel, and Fujitsu and was re-sponsible for advanced technology products fromconcept to delivery. He is currently a Professor inthe Department of Computer Science and Engineer-

ing, University of North Texas, Denton, TX, where he is also the FoundingDirector of the Network Security Laboratory, which aims to study the problemsand issues related to next-generation networks. He was a Visiting Professor atthe Massachusetts Institute of Technology (MIT), Cambridge, from Jan. 2011 toDec. 2011. His research interests include wireless networks, network security,security in voice and video over IP, and mobile and pervasive computing.

Srikanth Jonnada received the B.Tech. degree ininformation technology from JNT University, Hyder-abad, India, in 2005 and the M.S. degree in computerscience with a major in computer networking fromthe University of North Texas, Denton, in 2012.

His research interests include software de-velopment, mobile technologies, and bio-medicalapplications.

Shanti Thiyagaraja received the B.Eng. degree inelectrical and electronics engineering from the INTIInternational College of Penang, Penang, Malaysia,in 2010. She is currently working toward the Ph.D.degree at the Network Security Laboratory, ComputerScience and Engineering Department, University ofNorth Texas, Denton, TX.

Her research interests include signal process-ing, pattern recognition and mobile health careapplications.

Kalyan Pathapati Subbu received the B.E. degreein electronics and communication from the Univer-sity of Madras, Chennai, Tamil Nadu, India, in 2004,the Master’s degree in electrical engineering from theUniversity of North Carolina, Charlotte, NC, in 2007and the Ph.D. degree in computer science and engi-neering from the University of North Texas, Denton,TX.

He is currently a Postdoctoral Researcher atNanyang Technology University, Singapore. His re-search interests include mobile and pervasive com-

puting, pattern recognition, signal processing, and wireless sensor networks.