ROBUST EAR LOCATED HEART RATE MONITOR

L. Rossini, R. Vetter, C. Verjus, P. Theurillat, P. Renevey, M. Bertschi and J. Krauss

Swiss Center for Electronics and Microtechnology, Jaquet-Droz 1, Neuchˆatel, Switzerland

Keywords:

Ear located heart rate monitor, Strapless heart rate monitor, Ear clip, Earphone, Portable audio players.

Abstract:

We have developed a device for heart rate estimation with the optical sensing unit integrated in a classic

media player earphone. The sensing principle is based on an optical infrared measurement directly on the ear

lobe whereas the heart rate estimation is obtained utilizing robust model-based signal processing techniques

valid even for quasi periodic activities such as running. Nevertheless, the remaining problem related to these

strategies are statistically undetectable inter-beat during short term sporadic activities. In this paper, we present

a novel robust inter-beat discarding method based on an activity related modeling of the expected heart rate

dynamics that incorporates a simple cardiovascular model to reduce related inaccuracies in the heart rate

estimation. A validation protocol has been designed and 9 subjects were asked to carry out their daily normal

ofﬁce activities for a time length ranging from 1 to 2 hours. A global absolute relative error mean of 0.9%

between the estimated heart rate and a reference device and a sensitivity above 90% demonstrate encouraging

performances of the proposed device.

1 INTRODUCTION

The advent of portable audio players is continuously

increasing the popularity of exercising with music.

Besides enjoying their favorite songs, people may ﬁnd

themselves often more motivated toward the achieve-

ment of a given training load (Elliott et al., 2005;

Schie et al., 2008). However, to obtain a desired

training effect such as weight-loss or improvement of

the cardiovascularperformance, the exercising person

should observe and respect its personal heart rate tar-

get zones (Noakes, 2003). This requires supplemen-

tary equipment such as a heart rate monitor with its

associated chest-strap ECG sensing. Often, due to

the tightness of the strap, the heart rate monitor and

chest-strap sensor may be perceived by users as op-

pressive and may consequently diminish the exhila-

rating and motivating feeling provided by music. In

order to avoid this degradation, a device for heart rate

estimation based on a sensing unit directly located in

a classical audio player earphone has been presented

recently in (Celka et al., 2004; Verjus et al., 2003).

The sensing is based on an infrared measurement at

the ear cartilage whereas the signal processing is di-

rectly performed in the audio player unit where audi-

tory user feedback may be achieved. A critical aspect

of this approach is the signal processing. Adaptive

model-based enhancement of infrared signals is per-

formed to obtain a robust heart rate estimation even

for quasi periodic activities such as for example run-

ning. A remaining problem of this device is a reli-

able management of heart rate estimation during short

term sporadic activities. Indeed, in this case, model

based IR signal enhancement may not operate accu-

rately due to the convergencetimeof the enhancement

model (Haykin, 1991). Consequently, residual move-

ment related artifacts might induce erroneous inter-

beat detections that may not be detected with a sta-

tistical assessment (Vaseghi, 1996). In order to cope

with this problem, we have developed a reliable dis-

carding of erroneous inter-beat intervals based on an

activity related modeling of the expected heart rate

dynamics. The exploitation of previous biomedical

knowledge allowed us to develop a simple cardiovas-

cular model to reduce inaccuracies in the heart rate

estimations.

214

Rossini L., Vetter R., Verjus C., Theurillat P., Renevey P., Bertschi M. and Krauss J. (2009).

ROBUST EAR LOCATED HEART RATE MONITOR.

In Proceedings of the International Conference on Biomedical Electronics and Devices, pages 214-219

DOI: 10.5220/0001541002140219

 SciTePress

2 METHODS

2.1 Background

Optical probes for sensing biological tissue proper-

ties based on photoplestysmography(PPG) have been

widely used over the past years for the estimation of

cardiovascular parameters such as for example pulse

oximetry and heart rate (Webster, 1997), (Tremper

and Barker, 1989). Corruption of the PPG signal

arises from the inﬂuences of ambient light and subject

motion (Tremper and Barker, 1989) (Trivedi et al.,

1997). Processing of ambient light artifacts is not crit-

ical since the inﬂuence can be measured using multi-

plexing techniques and an artifact free PPG signal can

be restored using subtractive-type techniques (Trivedi

et al., 1997). Various methods for improving the PPG

technique during motion artifacts and low perfusion

of the tissue have been designed (Coetzee and Elg-

hazzawi, 2000). A very sound and robust approach

of motion artifacts removal in PPG measured sig-

nals has been recently addressed (Celka et al., 2004).

The parametric signal enhancement method exploits

the information contained in a motion reference sig-

nal generated by a two-dimensional accelerometer in

order to obtain a robust PPG heart rate estimation.

Very reliable heart rate estimations have been ob-

tained even under intense physical activity such as for

instance running. However, because of the slow con-

vergence of the enhancement algorithm, the heart rate

estimation may become erroneous during sporadic,

short, and transitory activities. The convergence time

of the enhancement algorithm is in the range of the

few seconds that are necessary to obtain sufﬁcient ac-

curacy of the enhancement parameters.

2.2 Sensor and Processing Device

In this paper we propose a fully integrated heart rate

measurement device with sensing located at the ear

and providing reliable estimates of the heart rate that

are robust against short sporadic or transitory activi-

ties. This system is based on infrared optical measure-

ment of the sub-cutaneous blood ﬂow by transillumi-

nation, together with an integrated two-dimensional

accelerometer (see Figure 1). The chosen optical

wavelength is 875 nm. The emitter is a light emit-

ting diode (LED) and the receiver is a photodetector

(PD). The light wave is sent trough the ear cartilage

and penetrates the skin and blood vessels to ﬁnally

reach the PD. The PD transforms the received light

intensity I(t) into a current that is then transformed

into a voltage. Subsequently, as shown in Figure 2, a

compensation of the ambient light is performed using

Figure 1: Ear located sensor device with portable process-

ing unit.

a subtractive multiplexing technique (Trivedi et al.,

1997).

The resulting signal IR(t) together with the two ana-

log acceleration signals a

(t) and a

(t) are then con-

ditioned through ampliﬁcation and a 2nd order But-

terworth bandpass ﬁlter between 0.5 Hz and 3.5 Hz.

Finally, an analog-to-digital (ADC) conversion is per-

formed on these signals at a sampling frequency of

20 Hz. Signals are processed in the portable unit and

the resulting heart rate estimations are displayed to

the user. A direct auditory user feedback through the

sensing earphone is an attractive option for a future

development.

Figure 2: The optical concept and electronic signal con-

ditioning of the proposed ear located heart rate estimation

device.

2.3 Physical Principle

The principle of the proposed method resides in in-

jecting an optical infrared (IR) signal at the surface

of the body tissue and measuring the resulting op-

tical signal. This signal propagates through the ear

tissue where it is subject to modiﬁcations due to re-

ﬂection, refraction, scattering and absorption. The

resulting signal is captured by one or multiple opti-

cal sensors distributed on the earphone. For the near

IR wavelength, the light propagation into the tissue

ROBUST EAR LOCATED HEART RATE MONITOR

215

is primarily governed by scattering and absorption

(Cheong et al., 1990). The Beer-Lambert equation

is generally used to describe the phenomenon of light

absorption in biological tissue (Coetzee and Elghaz-

zawi, 2000) relating the injected light I

to the output

light I

. However, motion artifacts affect the compo-

nents of the Beer-Lambert equation. Under this con-

ditions, the received intensity can be written in terms

of the major contributions:

(t) = I

(t) · γ

tissue

· γ

pulse

(t) · γ

motion

(t) (1)

where γ

tissue

is the static attenuation due to the tissue,

pulse

(t) is the pulsatile component due to variations

in the sub-cutaneous blood ﬂow, and γ

motion

(t) is the

contribution due to dynamic changes of the tissue in-

duced by movements of the head. The contribution

of γ

pulse

(t) in equation 1 is of pivotal interest for the

heart rate estimation. On the other hand, the time-

invariant term γ

tissue

is of no interest and can therefore

be removed using low-pass ﬁltering. Adaptive sig-

nal processing enhancement techniques of the signal

(t) to cope with long term harmonic contribution of

motion

(t) have been previously successfully presented

in (Celka et al., 2004). The technique developed dur-

ing the present activity addresses the problem of the

heart rate estimation during short and sporadic activi-

ties.

2.4 Proposed Algorithm

2.4.1 Concept

The concept of the proposed ear located heart rate

estimation device is shown in Figure 3. It mainly

consists of an enhancement of the motion corrupted

IR signals, an inter-beat interval (IBI) extraction on

the enhanced IR signal, a discarding of unreliable

IBIs and a ﬁnal estimation of the most likely heart

rate through histogram clustering. The enhance-

ment method that receives one part of its parameters

from the activity analysis has been presented in detail

in (Renevey et al., 2001). The histogram clustering is

based on a standard method of statistical signal pro-

cessing (Gersho and Gray, 1992). The new features

that we introduce address the optimal sensor place-

ment strategy, the IR pulse-amplitude reliability as-

sessment, the activity analysis and the estimation of

a prior of heart rate using a simple model of the car-

diovascular dynamics. This prior estimation of heart

rate and its associated conﬁdence intervals together

with the reliability index of the IR-pulse-amplitude

are then used to discard unreliable IBIs.

2.4.2 IR-Pulse-Amplitude Reliability

The reliability index of IR-pulse-amplitude describes

the likelihood that a given IBI has been extracted from

an IR signal with an amplitude that corresponds to

the expected amplitude of pulse contributions. It is

obtained by evaluating the probability that the instan-

taneous IR amplitude has been generated by a pro-

cess with mean µ

IR,ref

and standard deviation σ

IR,ref

which are continuously re-estimated during periods

without any movement activity.

Figure 3: The proposed heart rate estimation algorithm

based on motion artifact removal using accelerometer sig-

nals and discarding of unreliable inter-beat intervals using

activity related cardiovascular modeling.

2.4.3 Optimal Sensor Placement Strategy

Due to peripheral vasoconstriction and under-optimal

sensor placement, IR pulse contribution may be very

low. Therefore, we introduce an optimal sensor place-

ment procedure that evolves during the initialization

phase of the device:

• The signal amplitude is fed back to the user by

means of a quality index.

• As long as the quality index is below a given

threshold, the user is asked to adjust the sensor

placement to improve the signal quality.

We observed that this strategy leads to a sufﬁcient

quality of the IR pulse contribution. In about 2 %

of the cases only, the subject was unable to properly

adjust the device in an appropriate position. This pro-

cedure could also be processed fully automatically if

a multi-sensor approach is used.

2.4.4 Activity based Model Heart Rate Model

The proposed ear located sensor is based on IR sig-

nals and as such is prone to movement related arti-

facts. During transitory periods when the enhance-

BIODEVICES 2009 - International Conference on Biomedical Electronics and Devices

216

ment method cannot converge, these artifacts may in-

duce erroneous IBI detections. To discard such er-

roneous detections we improve the statistical robust-

ness of the proposed heart rate estimator by introduc-

ing a modeling of the expected heart rate based on

accelerometer measurements. The study and devel-

opment of cardiovascular models has attracted a wide

spreading interest in the scientiﬁc community(see (Le

et al., 2008) and references therein). In our appli-

cation, we are particularly interested in the descrip-

tion of the heart rate dynamics with respect to a given

input activity. Such an approach has been recently

proposed for cycling (Le et al., 2008). In particular,

the instantaneous heart rate is modeled as a weighted

function of its past value and the power generated by

the cyclist that is estimated using commercial devices.

Notice, however, that for activities like running and

walking the generated power output is not obviously

measurable. Nevertheless, during running and walk-

ing the exerciser’s power output is approximately re-

lated to his speed and a method for speed estimation

using core body located accelerometers has been re-

cently presented (Vetter et al., 2008). Even though

core body accelerometer measurements are not iden-

tical to ear located accelerometer measurement due to

slight head movements during each stride cycle, we

may assume that they are strongly related. We exploit

the method presented in (Vetter et al., 2008) to obtain

a rough measurement of the power output of a run-

ner P(n) as a function of the maximal eigenvalue of

stride-wise speed variations.

The heart rate model estimation HR

model

(n) at

sample n is a combination of the HR value before the

exercise HR

rest

(n) and a heart rate increase ∆HR(n)

model

(n) = HR

rest

(n) + ∆HR(n) (2)

The dynamic behavior of the heart rate variation un-

der physical effort is described by

∆HR(n) = α∆HR(n− 1) + (1− α)HR

ATR

(n) (3)

where α determines the temporal behavior of our

model. Speciﬁcally, we use α

act

when the activity is

increasing or stationary and α

recov

during a recovery

phase. For this model, the evolution of the heart rate

is a function of both its past values and the innovation

term HR

ATR

(n). The innovation term combines the

power output of an aerobic effort P(n) with the differ-

ence between the cardiac anaerobic threshold HR

and the heart rate at rest HR

rest

(n).

ATR

(n) = α

[HR

− HR

rest

(n)]P(n) (4)

The main goal of our model is to obtain a rough es-

timation of the heart rate in order to discard non-

plausible instantaneous estimations. In this sense,

to the various unknown parameters α

, α

act

, α

recov

and HR

of our model, we apply commonly utilized

sport physiological values (Noakes, 2003). These

values are then updated during the use of the device

so as to correspond to a user’s speciﬁc proﬁle.

2.4.5 HR Reliability

Sporadic motion artifacts augment the number of dis-

carded marginal values, reducing the reliability of the

estimated HR. For this reason, together with the esti-

mated HR, the user is provided with a HR reliability

index deﬁned as

rel

(n) =

IBI,∆T

(n)

∆T · µHR

∆T

(n)

(5)

where ∆T is the window length multiple of the sam-

pling time, N

IBI,∆T

(n) and µHR

∆T

(n) are respectively

the number of valid IBIs and the expected mean HR

at the sample n over the time span ∆T.

3 RESULTS

In order to validate the developed estimation tech-

nique, nine subjects were requested to follow an ad-

equate experimental protocol. Subjects were asked

to carry out their daily normal ofﬁce work for a to-

tal time length comprised between 1 to 2 hours and

to carry out 2 to 3 high intensity activities lasting ap-

proximately 10 to 20 seconds. In addition to the de-

veloped ear located heart rate device, subjects were

equipped with a commercial chest-strap based heart

rate monitor (POLAR, RS800) employed as a refer-

ence.

To illustrate the performances of the proposed algo-

rithm, we present quantitative and qualitative results

comparing the heart rate estimated by the proposed

method to the POLAR reference. The most com-

monly applied quantitative assessments are based on

the Mean Absolute Relative Error (MARE) and Mean

Absolute Error (MAE):

MARE = 100

∑

n=1



HR(n) − HR

polar

(n)



polar

(n)

(6)

MAE =

∑

n=1



HR(n) − HR

polar

(n)



(7)

Furthermore, we compute a mean reliability index

) associated to each measure. R

is obtained

by adding all the HR reliabilities greater than 0.5 and

normalizing by the measurement length N, where HR

ROBUST EAR LOCATED HEART RATE MONITOR

217

reliabilities have been previously deﬁned in equation

= 100

∑

n=1

(HR

rel

(n) > 0.5) (8)

The promising performances relative to the entire

database are presented in Table 1. Indeed, for the

whole database, the mean µ and standard deviation σ

associated to the MARE are 0.9 and 0.6 respectively.

It is important to observe that the objective of this

validation is to demonstrate the ability of our device

to accurately estimate the heart rate under baseline

resting conditions without concentrating on heart rate

variabilities (HRVs) (of the European Society of Car-

diology et al., 1996). For this reason, we ﬁltered the

heart rate provided by the POLAR to eliminate HRVs

above 0.04 Hz. Therefore, HRVs associated to the

sympathetic and parasympathetic nervous system are

not retained in the validation.

Table 1: Mean Absolute Relative error (MARE), Mean Ab-

solute Error (MAE), and reliability for 9 subjects in baseline

conditions with intermittent sporadic activities.

Subject MARE [%] MAE [bpm] R

[%]

1 0.4 0.3 98

2 0.3 0.6 95

3 1.2 1.0 92

4 1.2 0.8 94

5 0.6 0.8 94

6 1.4 1.0 90

7 0.7 0.5 99

8 2.8 2.4 93

9 1.6 0.9 93

µ 0.9 1.1 94

σ 0.6 0.8 2.8

Although the MARE and MAE provide an infor-

mation about the average estimation performance of

the developed device when compared to the POLAR

RS800, they both only partially describe the impres-

sion perceivedby the user. Indeed, the estimation per-

formance may be excellent for long periods despite

being erroneous for some short time intervals. There-

fore, to take into account this temporal variation, we

also apply an analysis of the speciﬁc threshold sensi-

tivity, namely, an assessment of the performance as

the percentage of time where the absolute value of

the error is lower than a given threshold κ. The spe-

ciﬁc threshold sensitivity of the algorithm is presented

in Table 2. The percentage of time where the error

in heart rate estimation is lower than the indicated

threshold κ highlights that low sensitivities are ob-

tained for a threshold of 1% and 1bpm respectively.

However, since the accuracy of the reference heart

rate monitor is about ±1%, the analysis may not be

Table 2: Percentage of time where the error in heart rate

estimation is lower than the indicated threshold κ .

% bpm

Subject 1 3 5 1 3 5

1 92 99 100 95 100 100

2 90 95 97 93 97 99

3 56 88 96 62 90 97

4 68 89 95 76 92 97

5 74 94 98 82 97 99

6 48 85 94 62 91 97

7 77 98 100 84 99 100

8 33 68 83 46 80 91

9 62 83 92 72 92 97

µ 81 90 95 75 93 97

σ 19 10 5 16 6 3

of high statistical relevance. A validation with such

an accuracy should be performed using medical refer-

ence devices. In contrast, sensitivities for a threshold

of 3% and 5% are above or equal to 90%.

Figure 4 illustrates qualitative performances for

the ﬁrst subject in the database. One can observe the

high accuracy of the proposed method. Notice that

at the beginning of the recording, the IR signal qual-

ity was insufﬁcient to provide a valid heart rate esti-

mation because the user was unable to properly ad-

just the ear located sensor. However, once the sen-

sor placement strategy is successfully achieved, heart

rate estimation converged as conﬁrmed by a MARE

of 0.4% and relative heart rate estimation error lower

than 1bpm for 99% of the time (see subject number 1

of Table 1 and Table 2).

In Figure 5 we depict the result relative to subject

8. A closer visual inspection on the data highlights

that there are mainly three segments where the results

of the POLAR and of our device diverged. To illus-

trate, we observe two segments at the beginning of the

recording where the heart rate estimated by POLAR

is about 170 bpm despite weak accelerometer signals

corroborating that the subject was in a resting posi-

tion. This unlikely estimation may be the result of

an incorrect manipulation of the chest strap reference

device such as for instance an insufﬁcient humidiﬁca-

tion. Finally, for the third segment at about 1300 sec-

onds where we notice a difference between the two

heart rate estimates there is an insufﬁcient reliability

rel(n)

of the heart rate provided by our device.

BIODEVICES 2009 - International Conference on Biomedical Electronics and Devices

218

Figure 4: Heart rate estimation for subject 1 in baseline

condition with sporadic intermittent activities using the pro-

posed algorithm.

Figure 5: Heart rate estimation for subject 8 in baseline con-

dition with sporadic intermittent activities.

4 CONCLUSIONS

In this paper we have presented an innovative ap-

proach for heart rate estimation using infrared opti-

cal measurements of the tissue at the ear lobe. The

sensing device integrated in a classical audio player

earphone improves the user’s comfort with respect to

commonly used chest-strap based heart rate monitors.

The challenge connected to statistically undetectable

inter-beat during short term sporadic activities has

been approached using a novel robust inter-beat dis-

carding method. The latter incorporates a simple car-

diovascular dynamic model to reduce related inaccu-

racies in the heart rate estimation. Nine subjects were

asked to participate in the validation protocol of the

proposed device. The validation resulted in positive

performances with an absolute relative error mean of

0.9% and a threshold sensitivity above 90% relative

to the chest-strap heart reference monitor.

REFERENCES

Celka, P., Verjus, C., Vetter, R., Renevey, P., and Neuman,

V. (2004). Motion resistant earphone locatedinfrared

based heart rate measurement device. In IASTED Con-

ference on Biomedical Engineering (BioMED 2004),

pages 582–585.

Cheong, W.-F., Prahl, S., and Welch, A. (1990). A review of

the optical properties of bilogical tissues. IEEE Jour-

nal Quantum Electronic, 26:2166–2185.

Coetzee, F. M. and Elghazzawi, Z. (2000). Noise-resistant

oximetry using a synthetic reference signal. IEEE

Trans. on Biomedical Engineering, 47(8):1018–1026.

Elliott, D., Carr, S., and Omre, D. (2005). The effect of mo-

tivational music on sub-maximal exercise. European

Journal of Sport Medicine, 5:97–106.

Gersho, A. and Gray, R. (1992). Vector Quantization and

Signal Compression. Kluwer Academic Publishers.

Haykin, S. (1991). Adaptive Filter Theory. Prentice Hall.

Le, A., Jaitner, T., Tobias, F., and Litz, L. (2008).

A dynamic heart rate prediction model for train-

ing optimization in cycling. In The Engineer-

ing of Sport(ISEA2008), volume 1, pages 425–433.

Springer.

Noakes, T. (2003). Lore of Running. Ed. Champaign.

of the European Society of Cardiology, T. F., the North

American Society of Pacing, and Electrophysiology

(1996). Heart rate variability - standards of measure-

ment, physiological interpretation, and clinical use.

Circ., 93:1043–1063.

Renevey, P., Vetter, R., Krauss, J., Celka, P., and De-

peursinge, Y. (2001). Wrist-located pulse detectionus-

ing ir signals, activity and nonlinear artifact cancella-

tion. In IEEE EMBS Conference.

Schie, N., A.Stuart, Becker, P., and Rogers, G. (2008). Ef-

fect of music on sub-maximal cycling. South African

Journal of Sport Medicine, 20:28–31.

Tremper, K. and Barker, S. (1989). Pulse oximetry. Anes-

thesiology, 70:98–108.

Trivedi, N., Ghouri, A., Shah, N., Lai, E., and Barker, S.

(1997). Effect of motion, ambient light, and hypoper-

fusion on pulse oximeter function. Journal of Clinical

Anesthesia, 9:179–183.

Vaseghi, S. (1996). Advanced signal processing and digital

noise reduction. Wiley Teubner.

Verjus, C., Vetter, R., Celka, P., and Renevey, P. (2003).

Equipement portable destine a la mesure et/ou la

surveillance de la frequence cardiaque. In European

Patent EP 1 374 763, US Patent 7175601.

Vetter, R., Onillon, E., and Bertschi, M. (2008). Estimation

of a runner’s speed based on chest-belt integrated iner-

tial sensors. In The Engineering of Sport(ISEA2008),

volume 1, pages 151–159. Springer.

Webster, J. (1997). Design of pulse oximeters. IoP Publish-

ing.

ROBUST EAR LOCATED HEART RATE MONITOR

219